close

Вход

Забыли?

вход по аккаунту

?

The impact of snow cover variability on snow water equivalentestimates derived from passive microwave brightness temperatures over a prairieenvironment

код для вставкиСкачать
THE IMPACT OF SNOW COVER VARIABILITY ON
SNOW WATER EQUIVALENT ESTIMATES
DERIVED FROM
PASSIVE MICROWAVE BRIGHTNESS TEMPERATURES
OVER A PRAIRIE ENVIRONMENT
A Thesis
Submitted to the Faculty of Graduate Studies and Research
In Partial Fulfillment of the Requirements
for the Degree of
Master of Science
in Geography
University of Regina
by
Kim Richard Turchenek
Regina, Saskatchewan
March, 2010
Copyright 2010: K.R. Turchenek
1*1
Library and Archives
Canada
Bibliotheque et
Archives Canada
Published Heritage
Branch
Direction du
Patrimoine de I'edition
395 Wellington Street
OttawaONK1A0N4
Canada
395, rue Wellington
Ottawa ON K1A 0N4
Canada
Your file Votre r6f6rence
ISBN: 978-0-494-65726-3
Our file Notre inference
ISBN: 978-0-494-65726-3
NOTICE:
AVIS:
The author has granted a nonexclusive license allowing Library and
Archives Canada to reproduce,
publish, archive, preserve, conserve,
communicate to the public by
telecommunication or on the Internet,
loan, distribute and sell theses
worldwide, for commercial or noncommercial purposes, in microform,
paper, electronic and/or any other
formats.
L'auteur a accorde une licence non exclusive
permettant a la Bibliotheque et Archives
Canada de reproduire, publier, archiver,
sauvegarder, conserver, transmettre au public
par telecommunication ou par I'lnternet, preter,
distribuer et vendre des theses partout dans le
monde, a des fins commerciales ou autres, sur
support microforme, papier, electronique et/ou
autres formats.
The author retains copyright
ownership and moral rights in this
thesis. Neither the thesis nor
substantial extracts from it may be
printed or otherwise reproduced
without the author's permission.
L'auteur conserve la propriete du droit d'auteur
et des droits moraux qui protege cette these. Ni
la these ni des extraits substantiels de celle-ci
ne doivent etre imprimes ou autrement
reproduits sans son autorisation.
In compliance with the Canadian
Privacy Act some supporting forms
may have been removed from this
thesis.
Conformement a la loi canadienne sur la
protection de la vie privee, quelques
formulaires secondaires ont ete enleves de
cette these.
While these forms may be included
in the document page count, their
removal does not represent any loss
of content from the thesis.
Bien que ces formulaires aient inclus dans
la pagination, il n'y aura aucun contenu
manquant.
1+1
Canada
UNIVERSITY OF REGINA
FACULTY OF GRADUATE STUDIES AND RESEARCH
SUPERVISORY AND EXAMINING COMMITTEE
Kim Richard Turchenek, candidate for the degree of Master of Science in Geography, has
presented a thesis titled, The Impact of Snow Cover Variability on Snow Water
Equivalent Estimates Derived from Passive Microwave Brightness Temperatures
Over a Prairie Environment, in an oral examination held on December 15, 2009. The
following committee members have found the thesis acceptable in form and content, and
that the candidate demonstrated satisfactory knowledge of the subject material.
External Examiner:
Dr. Yee-Chung Jin,
Faculty of Engineering and Applied Science
Supervisor:
Dr. Joseph M. Piwowar, Department of Geography
Committee Member:
Dr. Kyle Hodder, Department of Geography
Chair of Defense:
Dr. Phillip Hansen, Department of Philosophy and Classics
ABSTRACT
Considerable seasonal and inter-annual variation in the physical properties and extent of
snow cover pose problems for obtaining reliable estimates of quantities and
characteristics of snow cover from both conventional and satellite measurements
(Goodison and Walker, 1994; Goita et al, 2003). In spite of these challenges, the Climate
Research Branch of the Meteorological Service of Canada (MSC) has developed a suite
of algorithms to derive snow water equivalent (SWE) estimates from remotely sensed
passive microwave imagery (Goodison and Walker, 1994; Derksen et al., 2002a; Goita et
al., 2003). The MSC algorithms work particularly well over open prairie environments
under the assumption of large areas of consistent snow cover (Derksen et al., 2002a).
While studies have documented underestimation in passive microwave estimates of snow
extent in marginal areas when compared to optical satellite data (Derksen et al., 2003b),
the accuracy in SWE retrievals under variable and patchy snow conditions is not well
understood.
In an effort to better understand how a variable and patchy snow cover impacts remotely
sensed SWE retrievals, field-based experiments were conducted over patchy snow
covered areas in February 2005 and March 2008. A systematic sampling strategy was
developed over a 1600 square kilometre (km ) area in southern Saskatchewan near a
calibration/validation flight line used for algorithm development in the 1980s (Goodison
and Walker, 1994). Land covers found at the sampling sites included fallow and stubble
fields, pastures and shelter belts. A large number of sampling sites contained snow pack
layers that included one or more ice lenses.
This research verifies that the continuous snow cover assumption embedded in the MSC
passive microwave SWE algorithm does not produce acceptable results over a patchy
snow cover. Several in-situ observations that appear to play an important role in affecting
the satellite passive microwave data over a variable snow cover include the presence or
absence of an ice lens, the fractional snow covered area, snow depth and the ground
temperature. In an attempt to mitigate the impact of fractional snow cover on snow water
equivalent estimates, a weighted algorithm is proposed that applies the percentage of
snow cover over a remotely sensed footprint to the SWE estimate derived by the MSC
algorithm.
11
ACKNOWLEDGEMENTS
Funding for this project was granted by CRYSYS - CRYosphere SYStem in Canada, a
collaborative research initiative partly funded by Environment Canada's Meteorological
Service of Canada, as well as by the University of Regina's Faculty of Graduate Studies
and Research, and The Environmental Research and Response Applications (TERRA)
Lab at the University of Regina's Department of Geography.
The EASE-Grid brightness temperatures were obtained from MSC through the EOSDIS
National Snow and Ice Data Center Distributed Active Archive Center (NSIDC DAAC),
University of Colorado at Boulder.
Special thanks are extended to Dr. Joseph M. Piwowar (University of Regina) for
supervising this research project, Dr. Chris Derksen (Meteorological Service of Canada)
for support and feedback throughout the project, and to Natasha Neumann, Arvids Silis,
and Peter Toose (all from the Meteorological Service of Canada) for equipment and data
support.
Acknowledgements are also extended to Andre Piasta, Aaron Fedje, Kari Geller, Kathie
Legault, Mark Otterson, Greg Peterson, Susan Rever, Mauricio Jimenez Salazar, and
Michelle Yaskowich (all of the University of Regina) for their assistance with the
collection of field measurements and observations.
in
TABLE OF CONTENTS
ABSTRACT
i
ACKNOWLEDGEMENTS
iii
CHAPTER ONE - INTRODUCTION
1
1.1
Research Purpose and Objectives
2
1.2
Rationale
3
1.3
Previous Research
3
1.4
Thesis Structure
4
CHAPTER TWO - STUDY AREA
5
2.1
Study Area Location
5
2.2
Climate
6
CHAPTER THREE - SNOW PROPERTIES
8
3.1
Snow Properties
8
3.2
Remote Sensing of Snow Properties
9
3.2.1 Visible and Near Infrared Imagery
10
3.2.2 Thermal Infrared Imagery
12
3.2.3 Passive Microwave Imagery
12
3.3
Microwave Energy Interactions in Snow
14
3.4
Summary
16
CHAPTER FOUR - METHODS
17
4.1
Field Campaigns
17
4.2
Data Collection
17
4.2.1 In-situ Data Collection
19
4.3
Remote Sensing Data
22
4.4
Data Processing and Analysis
25
4.4.1 In-situ Data Processing
25
4.4.2 Remote Sensing Data Processing
26
4.4.3 Map Queries
29
4.5
Statistical Tests
29
CHAPTER FIVE - RESULTS AND DISCUSSIONS
32
5.1
Statistical Analyses
32
5.1.1 In-situ vs. Passive Microwave SWE Estimates
32
5.1.2 Linear Regression Models
34
5.2
Spatial Analysis Results
39
5.2.1 Fractional Snow Cover from Thematic Mapper Imagery
39
5.2.2 25 km SSM/I SWE Weighted by TM Snow Fractions
42
5.2.3 25 km AMSR-E SWE Weighted by TM Snow Fractions
44
5.2.4 12.5 km AMSR-E SWE Weighted by TM Snow Fractions
46
5.2.5 Fractional Snow Cover from MODIS Imagery
59
5.2.6 25 km SSM/I SWE Weighted by MODIS Snow Fractions
62
5.2.7 25 km AMSR-E SWE Weighted by MODIS Snow Fractions
72
5.2.8 12.5 km AMSR-E SWE Weighted by MODIS Snow Fractions ...78
IV
TABLE OF CONTENTS
5.3
Spatial Analysis Validation Results
5.3.1 Fractional Snow Cover Validation from TM Imagery
5.3.2 Fractional Snow Cover Validation from MODIS Imagery
5.4
Results Discussion
5.4.1 Fractional Snow Cover Derived from TM Imagery
5.4.2 Fractional Snow Cover Derived from MODIS Imagery
CHAPTER SIX - CONCLUSIONS
6.1
In-situ vs. Passive Microwave SWE Estimates
6.2
MSC Algorithm Performance over a Partial Snow Cover
6.3
Manifestation of Spatial Variability in SWE Estimates
6.4
Degree of Fractional Snow Cover Influence on SWE Estimates
6.5
Summary
6.6
Future Research
REFERENCES
90
91
94
97
97
103
110
110
112
112
114
116
117
119
LIST of TABLES
Table 1
Table 2
Table 3
Table 4
Table 5
Table 6
Table 7
Table 8
Table 9:
Table 10:
Table 11:
Table 12:
Table 13
Table 14
Table 15
Table 16
Table 17
Table 18
Snow properties for environmental monitoring
Spectral regions and sensors used in snow research
Measured snow properties used in linear regression models
Optical imagery used within this study
Passive microwave data used within this study
Number of equivalent Snow-Only coincident sites
Number of equivalent Actual-Conditions coincident sites
Mean differences in SWE between remote sensing estimates and in-situ
measurements
Actual-Conditions SWE, depth and density measurements by land cover
Example co-efficient table output
Regression results between Swath AMSR-E SWE estimates and in-situ
observations
Regression results between remotely sensed SWE estimates and in-situ
observations
25 km SSM/I SWE estimates integrated with TM imagery
25 km AMSR-E SWE estimates integrated with TM imagery
12.5 km AMSR-E SWE estimates integrated with TM imagery
25 km SSM/I SWE estimates integrated with MODIS imagery
25 km AMSR-E SWE estimates integrated with MODIS imagery
12.5 km AMSR-E SWE estimates integrated with MODIS imagery
v
8
9
22
23
24
32
33
33
...34
35
36
38
99
99
101
105
105
107
LIST of FIGURES
Figure 2.1:
Figure 2.2(a):
Figure 2.2(b):
Figure 4.1(a):
Figure 4.1(b):
Figure 4.1(c):
Figure 4.1(d):
Figure 4.2:
Figure 4.3:
Figure 4.4:
Figure 4.5:
Figure 5.1:
Figure 5.2:
Figure 5.3:
Figure 5.4:
Figure 5.5:
Figure 5.6:
Figure 5.7:
Figure 5.8:
Figure 5.9:
Figure 5.10:
Figure 5.11:
Figure 5.12:
Figure 5.13:
Figure 5.14:
Figure 5.15:
Figure 5.16:
Figure 5.17:
Figure 5.18:
Study area
Open, fallow field
Stubble, snow-trapping field
Pasture
Shelter belt
Fallow
field
Stubble
field
Systematic sampling strategy
Pixel centroids
Thiessen polygon footprints
Regression analyses flow chart
Landsat TM 2005 Standard False-Colour Composite
Landsat TM 2005 iNDSI Snow Classification
Landsat TM 2008 Standard False-Colour Composite
Landsat TM 2008 iNDSI Snow Classification
AMSR-E and SSM/I pixels on Landsat TM 2005 Standard
False-Colour Composite background
SSM/I 25 km TM Fractional Snow Cover for February 21-23, 2005
SSM/I 25 km SWE vs. TM-Weighted SWE graph for
February 21-23, 2005
AMSR-E 25 km TM Fractional Snow Cover for
February 21-23, 2005
AMSR-E 25 km SWE vs. TM-Weighted SWE graph for
February 21-23, 2005
AMSR-E 12.5 km Pixel 1 TM Fractional Snow Cover for
February 22, 2005
AMSR-E 12.5 km Pixel 1 SWE vs. TM-Weighted SWE graph for
February 21-23, 2005
AMSR-E 12.5 km Pixel 2 TM Fractional Snow Cover for
February 23, 2005
AMSR-E 12.5 km Pixel 2 SWE vs. TM-Weighted SWE graph for
February 21 and 23, 2005
AMSR-E 12.5 km Pixel 3 TM Fractional Snow Cover for
February 22, 2005
AMSR-E 12.5 km Pixel 3 SWE vs. TM-Weighted SWE graph for
February 21-23, 2005
AMSR-E 12.5 km Pixel 4 TM Fractional Snow Cover for
February 23, 2005
AMSR-E 12.5 km Pixel 4 SWE vs. TM-Weighted SWE graph for
February 21 and 23, 2005
AMSR-E 12.5 km Pixel 5 TM Fractional Snow Cover for
February 23, 2005
vi
5
7
7
18
18
18
18
19
28
28
30
39
39
40
40
41
43
44
45
46
47
48
49
50
51
52
53
54
56
LIST of FIGURES
Figure 5.19:
Figure 5.20:
Figure 5.21:
Figure 5.22:
Figure 5.23:
Figure 5.24:
Figure 5.25:
Figure 5.26:
Figure 5.27:
Figure 5.28:
Figure 5.29:
Figure 5.30:
Figure 5.31:
Figure 5.32:
Figure 5.33:
Figure 5.34:
Figure 5.35:
Figure 5.36:
Figure 5.37:
Figure 5.38:
Figure 5.39:
AMSR-E 12.5 km Pixel 5 SWE vs. TM-Weighted SWE graph for
February 23, 2005
AMSR-E 12.5 km Pixel 6 TM Fractional Snow Cover for
February 22, 2005
AMSR-E 12.5 km Pixel 6 SWE vs. TM-Weighted SWE graph for
February 22 and 23, 2005
AMSR-E and SSM/I pixels on MODIS Fractional Snow Cover
background
Coincident SSM/I 25 km Pixel 1 MODIS Fractional Snow Cover
for February 21, 2005 P.M
SSM/I 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph for
February 21, 2005 P.M
Coincident SSM/I 25 km Pixel 1 MODIS Fractional Snow Cover
for February 22, 2005 A.M
SSM/I 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph for
February 22, 2005 A.M
Coincident SSM/I 25 km Pixel 1 MODIS Fractional Snow Cover
for February 22, 2005 P.M
SSM/I 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph for
February 22, 2005 P.M
Coincident SSM/I 25 km Pixel 1 MODIS Fractional Snow Cover
for February 23, 2005 A.M
SSM/I 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph for
February 23, 2005 A.M
Coincident SSM/I 25 km Pixel 1 MODIS Fractional Snow Cover
for February 23, 2005 P.M
SSM/I 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph for
February 23, 2005 P.M
Coincident AMSR-E 25 km Pixel 1 MODIS Fractional Snow Cover
for February 21, 2005
AMSR-E 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph for
February 21, 2005
Coincident AMSR-E 25 km Pixel 1 MODIS Fractional Snow Cover
for February 22, 2005
AMSR-E 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph for
February 22, 2005
Coincident AMSR-E 25 km Pixel 1 MODIS Fractional Snow Cover
for February 23, 2005
AMSR-E 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph for
February 23, 2005
Coincident AMSR-E 12.5 km Pixel 1 MODIS Fractional Snow
Cover for February 23, 2005
vn
57
58
59
61
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
LIST of FIGURES
Figure 5.40:
Figure 5.41:
Figure 5.42:
Figure 5.43:
Figure 5.44:
Figure 5.45:
Figure 5.46:
Figure 5.47:
Figure 5.48:
Figure 5.49:
Figure 5.50:
Figure 5.51:
Figure 5.52:
Figure 5.53:
Figure 5.54:
Figure 5.55:
Figure 5.56:
Figure 5.57:
Figure 5.58:
AMSR-E 12.5 km Pixel 1 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005
Coincident AMSR-E 12.5 km Pixel 2 MODIS Fractional Snow
Cover for February 23, 2005
AMSR-E 12.5 km Pixel 2 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005
Coincident AMSR-E 12.5 km Pixel 3 MODIS Fractional Snow
Cover for February 23, 2005
AMSR-E 12.5 km Pixel 3 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005
Coincident AMSR-E 12.5 km Pixel 4 MODIS Fractional Snow
Cover for February 23, 2005
AMSR-E 12.5 km Pixel 4 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005
Coincident AMSR-E 12.5 km Pixel 5 MODIS Fractional Snow
Cover for February 23, 2005
AMSR-E 12.5 km Pixel 5 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005
Coincident AMSR-E 12.5 km Pixel 6 MODIS Fractional Snow
Cover for February 23, 2005
AMSR-E 12.5 km Pixel 6 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005
AMSR-E 12.5 km Pixel 1 TM Fractional Snow Cover for
March 2, 2008
AMSR-E 12.5 km Pixel 1 SWE vs. TM-Weighted SWE graph for
March 2, 2008
AMSR-E 12.5 km Pixel 2 TM Fractional Snow Cover for
March 2, 2008
AMSR-E 12.5 km Pixel 2 SWE vs. TM-Weighted SWE graph for
March 2, 2008
Coincident AMSR-E 12.5 km Pixel 1 MODIS Fractional Snow
Cover for March 2, 2008
AMSR-E 12.5 km Pixel 1 SWE vs. MODIS-Weighted SWE graph
for March 2, 2008
Coincident AMSR-E 12.5 km Pixel 2 MODIS Fractional Snow
Cover for March 2, 2008
AMSR-E 12.5 km Pixel 2 SWE vs. MODIS-Weighted SWE graph
for March 2, 2008
,
vin
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
95
96
97
LIST of ABBREVIATIONS
AC
AMSR-E
AVHRR
DMSP
EASE.Grid
EOS
ESC
ETM
GHz
GIS
GPS
iNDSI
IR
MODIS
MSC
NAD
NDSII
NDSI
NOAA
NSIDC
SMMR
SO
SPSS
SSM/I
SWE
SWEW
TB
TM
USGS
UTM
VNIR
Actual Conditions
Advanced Microwave Scanning Radiometer for EOS
Advanced Very High Resolution Radiometer
Defense Meteorological Satellite Program
Equal Area Scalable Earth Grid
Earth Observing System
Eastern Snow Conference
Enhanced Thematic Mapper
Gigahertz
Geographic Information System
Global Positioning System
Inverse Normalized Difference Snow Index
Infrared
Moderate Resolution Imaging Spectroradiometer
Meteorological Service of Canada
North American Datum
Normalized Difference Snow and Ice Index
Normalized Difference Snow Index
National Oceanic and Atmospheric Administration
National Snow and Ice Data Center
Scanning Multichannel Microwave Radiometer
Snow-Only
Statistical Package for the Social Sciences
Special Sensor Microwave Imager
Snow Water Equivalent
Weighted Snow Water Equivalent
Brightness Temperature
Thematic Mapper
United States Geological Survey
Universal Transverse Mercator
Visible and Near Infrared
IX
CHAPTER ONE - INTRODUCTION
Accurate measurements of snow properties over differing temporal and spatial scales
allow improvements in the understanding of how snow contributes to the local and global
environment (Goodison and Walker, 1995). Since snow cover is an integrator of
precipitation and temperature it may also be considered an indicator of climate change
(Goodison and Walker, 1993; 1995). Furthermore, snow contributes to the water supply
and its availability for agriculture. As such, a change in water supply over semi-arid
regions may result in serious social and economic impacts (Brown et ah, 2000).
The Canadian Prairies are part of a semi-arid region in which basic physical and chemical
processes are affected by the water produced from snowmelt. Physically, dry surface soils
are prone to erosion processes, and chemically, these soils must be recharged with water
to retain nutrients for agricultural crop growth (Laycock, 1972; Goodison, 1995). Further,
the social and economic structures of this region are vulnerable to the receipt of an
adequate winter snowfall to contribute to its water supply.
Snow cover also impacts the prairies biologically and environmentally. Interactions
between living organisms and their physical environment are strong indicators of the state
of the prairie ecosystem. These interactions work both ways: when organisms are
unhealthy, so too is their physical environment; likewise, when the physical environment
is unhealthy, so too are its organisms. Therefore, investigation and monitoring of snow
cover is important in understanding the processes affected by snow properties.
Due to the large spatial extent of the Canadian Prairies, it is humanly and economically
impractical to gather spatially intensive in-situ snow measurements that cover the entire
region. Previous research has found however that remote sensing data are useful
resources for measuring and monitoring properties of snow from regional to global scales
(Robinson, 1993; Goodison and Walker, 1995). Fortunately, researchers have established
a relationship between a snowpack's snow water equivalent and electromagnetic energy
emissions at microwave frequencies. For example, the Climate Research Branch of the
Meteorological Service of Canada (MSC) has developed a suite of algorithms to derive
snow water equivalent (SWE) estimates from passive microwave imagery (Goodison and
Walker, 1995; Walker and Goodison, 2000; Derksen et al, 2002a; Goita et al, 2003).
The SWE algorithms work particularly well over open prairie environments (Goodison
and Walker, 1995; Walker and Goodison, 2000) under the assumption that the snowpack
remains constant over large areas. Derksen et al. (2002b) however, question the validity
of this assumption and identify the seasonal variability in a snowpack's physical
properties as an algorithm limitation. Thus, there is a need to understand the impact of
snow cover variability on SWE estimates derived from remote sensing imagery.
1.1
Research Purpose and Objectives
The purpose of this thesis was to investigate the impact of snow cover variability on
SWE estimates derived from remote sensing imagery over an agricultural area located in
southern Saskatchewan. The study's main objectives were to:
i. determine how to best represent ground-truth data for a partial snow cover;
2
ii. determine how well the MSC prairie SWE algorithm performed over a partial
snow cover;
iii. document how spatial variations observed in a partially snow covered area
were manifest in passive microwave estimates;
iv. discover the degree of fractional snow cover that was required to influence a
passive microwave - derived SWE estimate at different spatial resolutions.
1.2
Rationale
This research is expected to aid in algorithm tuning to enhance the application of remote
sensing data for understanding the impacts of snow cover variability on the hydrological,
environmental, economic and social structures of Canada's prairie region.
1.3
Previous Research
Previous research (Goodison and Walker, 1993; Derksen et.al, 2002a) has been driven by
the need to develop algorithms for deriving estimates of key snow properties that are
thought to make major contributions to flood and drought monitoring and/or climate
change analyses. For example, MSC's SWE algorithms were originally empirically
derived from linear regressions that factor in the slope and intercept of the best-fit
regression line between ground measurements and remotely sensed estimates (Goodison
and Walker, 1994; Goita et al, 1997; Derksen et ah, 2002a; 2003a).
Research since then has examined the relationship between SWE and the electromagnetic
energy emissions obtained at microwave frequencies. Currently, the Meteorological
Service of Canada employs a suite of linear algorithms to retrieve SWE estimates from
3
passive microwave sensors. The MSC algorithms vary according to land cover, with
different coefficients used for open prairie, deciduous forest, coniferous forest, and sparse
forest (Goita et al, 1997; Singh and Gan, 2000; Derksen et al., 2003a; 2003b)
1.4
Thesis Structure
The remainder of this thesis is structured in the following order: The next chapter
introduces and describes the area under study. Microwave remote sensing of snow cover
is discussed in Chapter Three. Chapter Four describes the methodologies used to acquire
in-situ measurements and perform spatial analyses, and also introduces the coincident
remote sensing data. Chapter Five presents and discusses results of the analyses. Chapter
Six provides conclusions of the thesis and a discussion on future research.
4
CHAPTER TWO - STUDY AREA
2.1
Study Area Location
The study area is located approximately 100 kilometres (km) south of Regina,
Saskatchewan near the town of Radville and the villages of Pangman and Ceylon (Figure
2.1). This area was selected because it coincides with an area previously used by
Environment Canada for snow cover research (Goodison and Walker, 1994). In addition,
the snow cover typical of this area has been found to have regions of both patchy and
complete snow-cover (Goodison and Walker, 1994; Turchenek, 2004). The study area
covered over 1600 km to ensure complete spatial coverage of at least one pixel of the
coarsest remote sensing data collected for this study.
Figure 2.1: Study area
5
The Missouri Coteau, a bedrock escarpment featuring ice-shoved hills (Thraves et al.,
2007) cuts across the southwest portion of the study area, forming a low, rolling
topography. Although the natural vegetation in this area consists of short grasses, a large
portion of the land is under agricultural production, where wheat and other grains are
farmed. Pockets of trees and shrubs create shelter belts in areas with higher moisture
supply.
2.2
Climate
Relatively short warm summers and long cold winters are characteristic of
Saskatchewan's prairies (Hare and Thomas, 1979). With few perennial streams, much of
the region's water supply comes in the form of precipitation. However, the annual
precipitation in the region is relatively low, and evapotranspiration usually exceeds the
annual precipitation, creating an average water deficit by middle to late summer
(Laycock, 1972). Most annual precipitation falls in the summer, while February is usually
the driest month (Hare and Thomas, 1979).
The winter season provides relatively low amounts of snow. Extended periods of cold,
clear weather are interrupted by occasional blizzards with gusting winds. Warming
periods are frequent in the early and late winter (Laycock, 1972; Hare and Thomas, 1979;
Walker et al., 1995). Wind re-distributes the snow cover by removing snow from one
area and depositing it in another. Re-distributed snow collects in areas with vegetated
landcover, such as shelter belts and fields left with crop stubble, where there are barriers
to re-distribution. Figures 2.2(a) and 2.2(b) are photographs of two of the sampling sites
from this study that illustrate the differences in the amount of snow cover between
6
vegetated and non-vegetated land covers. Figure 2.2(a) shows an open, fallow field that
had no barriers for the collection of re-distributed snow, while Figure 2.2(b) shows an
agricultural field left with crop stubble for the collection of re-distributed snow.
Figure 2.2(a): Open, fallow
(original in colour)
field
Figure 2.2(b): Stubble, snow-trapping field
(original in colour)
Warming periods also impact the snow cover through freeze-thaw processes (Laycock,
1972; Walker et al., 1995). As the air and ground temperatures rise, the snow pack melts.
When the snow pack re-freezes, it becomes denser and shallower. Thus, along with a
change in vegetation, weather systems can also impart considerable variability in the
depth and density of a snow pack.
7
CHAPTER THREE - SNOW PROPERTIES
3.1
Snow Properties
A snow pack can make significant contributions to groundwater supplies, flood events
and droughts. Inter-annual variations in snow cover are also considered to be useful
indicators of climate change (Goodison and Walker, 1993). Key properties of snow that
are important to measure are its areal extent, SWE, depth, density, wet/dry state, melt
stage and surface temperature (Table 1).
Table 1: Snow properties for environmental monitoring (after Walker et al., 1995)
Snow Property
Contribution
Areal Extent
Affects regional hydrologic cycles by storing and subsequently releasing
water through melt periods, potentially creating flood or drought
conditions.
Impacts Earth's energy balance by reflecting solar radiation, contributing
to increasing temperatures when the extent of snow cover diminishes.
Water Equivalent
Snow Water Equivalent is the amount of water stored in a snowpack that
is available for release upon melting, and is thought to be the most
important snow property for hydrologic studies such as flood and drought
monitoring.
Is a function of the depth and density of a snowpack, where the
snowpack's water equivalent is calculated by multiplying the snowpack's
depth by its density.
Depth
Determining factor of a snowpack's water equivalent.
Density
Determining factor of a snowpack's water equivalent.
Wet/Dry State
Indicator of snowpack density. Wet snow indicates the snow grains are
coated with water, which increases the density of the snowpack.
Melt Stage
Indicator of areal extent, wet/dry state, and snowpack density.
Surface Temperature
Indicator of snowpack density, wet/dry state, and melt stage.
8
3.2
Remote Sensing of Snow Properties
Remote sensing has been used for many years in snow monitoring research, and
numerous sensors have been used to monitor various snow properties. Table 2 lists the
sensors that have been used to monitor various properties of snow, as well as each
sensor's spectral region and wavelength or frequency.
The capabilities and limitations of remote sensing data for measuring and monitoring
snow properties are described, by spectral region, below.
Table 2: Spectral regions and sensors used in snow research
Wavelength
a>
x
W
Spectral Region
Sensor
or
0
L.
S
cs
u
a.
V
en
t/2
a.
5
«3
73
9>
H
s-
ETM/TM (Landsat)
Frequency
0.45 - 2.35 \ua
*
v
AVHRR(NOAA)
0 . 5 8 - 1.10 |im
V
V
MODIS (Terra/Aqua)
0.62 - 2.15 n m
V
Thermal IR
ETM/TM (Landsat)
1 0 . 4 - 1 2 . 5 (am
Passive M W
SSM/I (DMSP)
19,37, 85 GHz
^/''83UMZ
HV
HV
•
•
•
•
SMMR (Nimbus-7)
18, 37
GHz
]f'3lKjtiz
HV
HV
•
•/
•
•
AMSR-E(Aqua)
19,37 GHz
•
•
•
•/
Attexlnc.
(Airborne Passive)
19,37,85 GHz
HV
^
•
•
•
Visible & Near IR
9
V
3.2.1
Visible and Near Infrared Imagery
Visible and near infrared (VNIR) sensors commonly used to estimate snow properties
include the National Oceanic and Atmospheric Administration's (NOAA) Advanced
Very High Resolution Radiometer (AVHRR), NASA's Earth Observing System (EOS)
Enhanced Thematic Mapper (ETM)/Thematic Mapper (TM) aboard several Landsat
platforms, and the Moderate Resolution Imaging Spectroradiometer (MODIS) aboard
both the Terra and Aqua EOS platforms.
Data collected from optical sensors, such as Landsat's Thematic Mapper, are not useful
for deriving depth or SWE estimates (Goodison and Walker, 1993). However, optical
sensors are useful for monitoring the areal extent of snow cover (Goodison and Walker,
1993; Robinson, 1993; Bussieres et al., 2002). For example, Landsat ETM/TM data have
been used to determine snow cover extent on local scales (Rees, 2006). The extent is
determined using a Normalized Difference Snow Index (NDSI). The standard NDSI
algorithm (Rees, 2006) applied to ETM/TM data is defined as:
NDSI = ( r 2 - r 5 ) / ( r 2 + r5)
(3.1)
where, r2 and rs represent ETM/TM bands 2 and 5, respectively. The index outputs cell
values between 0 and 1, and it has been found that a value above 0.4 is assumed to
indicate the presence of snow, although, this threshold value may vary by season (Rees,
2006). The threshold is determined for local areas by visually comparing the NDSI output
image to a multispectral composite image.
10
While the ETM/TM sensors have been found useful for determining snow cover extent
over local scales, the AVHRR and MODIS sensors are capable of monitoring
hemispheric and regional extent of snow cover (Goodison and Walker, 1993; Robinson,
1993; Bussieres et al, 2002). Data obtained from AVHRR and MODIS are appropriate
for monitoring snow cover over larger spatial areas because their spatial resolutions of 1
km and 500 m, respectively, assist in reducing data volumes. While AVHRR and MODIS
data have both been used to map areal extent of snow cover, Bussieres et al. (2002)
recommend that MODIS daily images be used instead of AVHRR weekly composites
since the latter provide inaccurate data when large weather systems affect the areal extent
of snow cover. As well as having a higher spatial resolution than AVHRR, MODIS also
has a greater number of spectral bands (Rees, 2006). Spectral bands 2, 4 and 6 on
MODIS have been used to determine the presence of snow, and can be used to calculate
the NDSI. The MODIS NDSI tends to overestimate snow cover over patchy snow
covered areas because mixtures of land cover types are likely to be contained within the
low spatial resolution pixels (Rees, 2006).
Another snow property capable of being derived from VNIR data is melt stage. Winther
(1993) proposes that histogram equalization techniques on Landsat TM Band 5 data are
capable of discriminating melting phenomena, while De Abreu and LeDrew (1997) show
that a Normalized Difference Snow and Ice Index (NDSII) obtained from AVHRR
Channels 1 and 2 has potential for monitoring the melt stage of terrestrial snow.
While VNIR sensors are capable of monitoring areal extent and melt stage snow
properties, limitations of all visible sensors are their inabilities to obtain data during
11
darkness and under cloud cover. Optical sensors such as Landsat's ETM/TM and
NOAA's AVHRR are also incapable of obtaining data from within or below the surface,
restricting their usefulness for deriving depth and SWE measurements (Goodison and
Walker, 1993).
3.2.2
Thermal Infrared Imagery
Although thermal IR data have been available from Landsat's TM since 1982 and
NOAA's AVHRR since 1987, there has not been much research performed on snow
properties with these data. TM band 6 has been found capable of measuring surface
temperatures (Rees, 2006), but Winther (1993) found derived surface temperature
estimates were too low.
The thermal IR channels aboard the Landsat and NOAA platforms are capable of
obtaining data in darkness conditions, which overcomes a major limitation of optical
sensors.
3.2.3
Passive Microwave Imagery
Many studies have involved the use of passive microwave data to monitor snow
properties such as areal extent, depth, SWE, and wet/dry state (Goodison and Walker,
1993; Sokol, et al., 1999; Walker et al, 1995). The Scanning Multichannel Microwave
Radiometer (SMMR) carried by the Nimbus-7 satellite and the Special Sensor
Microwave Imager (SSM/I) carried on the U.S. Defense Meteorological Satellite
Program's (DMSP) platforms are the longest-used sensors in passive microwave snow
monitoring research. Other passive microwave sensors include the Meteorological
12
Service of Canada's (MSC) airborne radiometer, and the Advanced Microwave Scanning
Radiometer for EOS (AMSR-E) aboard the EOS Aqua satellite.
The SMMR operated from 1978 to 1987 and although it was capable of monitoring the
areal extent of snow cover, it was limited by its relatively narrow swath width of 780 km
(Walker et al., 1995). As a result, it would take three to five days to collect enough
imagery to cover a broad region (Goodison and Walker, 1993). This was problematic
because large weather systems could significantly affect the snow cover's areal extent
part way through the sensor's acquisition cycle. Similarly, the SMMR's narrow swath
width also limited its capabilities of estimating SWE values and wet/dry state. On the
other hand, the successor to the SMMR, the SSM/I, has a swath width of 1400 km and is
capable of providing daily coverage (Goodison and Walker, 1993), which overcomes
these limitations.
In spite of the swath-width differences between the SMMR and the SSM/I, their spectral
properties are similar enough to enable their data to be combined for the purpose of
creating longer time-series, which are critical components for climate change research.
Derksen et al. (2003b) combined SMMR and SSM/I data to create a time-series that
extended from 1978 to 1999. The more recent AMSR-E sensor operates in similar
frequencies to SSM/I and SMMR (Imaoka et al., 2002; Shibata, 2002) so the integration
of AMSR-E data with SSM/I and SMMR data seems feasible for creating longer timeseries. Research where data from all three sensors has been used is just starting to be
published (e.g., Comiso and Nishio, 2008).
13
MSC's airborne passive microwave radiometer has been used in research studies since
1996 (Walker et al., 2002). The airborne radiometer is installed on a Twin Otter aircraft
and used in field campaigns to develop and validate algorithms for the monitoring of
climate variable research.
Remote sensing in the passive microwave spectral region has the advantage of being able
to obtain data in darkness and under all weather conditions (Walker et al., 1995). Passive
microwave remote sensing is limited, however, by the course spatial resolutions of the
sensors (Singh and Gan, 2000) and the corruption of the passive microwave signal when
liquid water is present within a snowpack (Walker and Goodison, 1993). This makes it
difficult to derive accurate SWE estimates from a melting snowpack and accurate
estimates of snow cover extent (Walker et al., 1995). However, it has been suggested that
with careful daily monitoring, the limitation of accurately deriving areal extent can be
overcome in open prairie regions (Walker et al., 1995).
While much research has been performed on areal extent of snow cover with passive
microwave data, the most intensive research with these data has focused on algorithms
used to derive SWE estimates for validation of ground-truth observations (Derksen et al.,
2002a; Goita et al, 1997; Goodison and Walker, 1994).
3.3
Microwave Energy Interactions in Snow
Microwave radiation is naturally emitted everywhere on Earth. Its measurable intensity
varies from place to place based on soil types, land covers, snowpack characteristics, and
other variables (Goita et al., 1997; Sokol et ah, 1999). At microwave frequencies above
14
15 gigahertz (GHz), the emitted radiation is scattered by snow particles as it passes
through the snowpack (Goita et al., 1997). An increase in the snowpack depth or grain
size results in an increase in scattering and subsequent lower microwave brightness
temperatures when measured above the surface (Goita et al., 1997). Wet and/or dense
snowpacks however, decrease the amount of scattering and produce near blackbody
emissions (Sokol et al., 1999). Complex snowpacks, containing numerous layers at
different densities, selectively influence microwave radiation, producing inconsistent
measurements (Sokol et al., 1999).
Radiation recorded by a microwave sensor is expressed as a brightness temperature (TB)
in Kelvin units (Goita et al., 1997; Sokol et al., 1999). One parameter commonly derived
from remotely sensed brightness temperatures is snow water equivalent, which is the
amount of water stored in a snowpack that is available upon melt.
The algorithms that have been developed to estimate SWE from passive microwave
measurements vary according to the type of land cover (Goita et al., 1997; Singh and
Gan, 2000). For example, the algorithm used by the Meteorological Service of Canada
(MSC) to derive SWE over the open prairie is a vertically (v) polarized TB gradient ratio
(Goodison and Walker, 1994; Derksen et al, 2003a) defined as:
SWE (mm) = -20.7 - (37v - 19v) * 2.59
(3.2)
Variables 37v and 19v are the brightness temperatures acquired from vertically polarized
frequencies of 37 and 19 GHz, while the coefficients 2.59 and -20.7 are the slope and
intercept of the best-fit regression line found between ground and airborne brightness
15
temperatures (Goodison, 1989). Numerous studies have found SWE estimates derived
from this algorithm to be within +/- 10-20 millimetres (mm) of in-situ observations
(Goodison and Walker, 1994; Derksen et al, 2002a; 2003a; 2003b), however, the MSC
algorithm assumes a complete snow cover and the effects of patchy snow cover on SWE
estimates are not well understood.
3.4
Summary
Previous research has shown that:
i.
data derived from the visible and near IR regions of the electromagnetic
spectrum are capable of monitoring the areal extent of snow cover and melt
stage, although patchy snow covered areas tend to be overestimated;
ii.
various spectral bands from various optical spaceborne platforms can be
applied to a Normalized Difference Snow Index to identify snow covered
areas;
iii. data derived from the thermal IR spectral region are deemed too low
to accurately derive surface temperatures;
iv. data derived from passive microwave sensors are capable of providing
accurate estimates of areal extent of dry snow cover, depth, SWE, and
wet/dry state;
v.
passive microwave derived SWE estimates are accurate to within +/- 10-20
mm of ground-truth observations;
vi. time-series can be extended by integrating data obtained from sensors
operating in similar frequencies or wavelengths.
16
CHAPTER FOUR - METHODS
4.1
Field Campaigns
To help understand the relationship between patchy snow cover and remotely sensed
SWE estimates, field campaigns to measure SWE under highly variable snowpack
conditions were conducted in late February 2005, with a follow-up validation campaign
in early March 2008. The timing of the field campaigns was crucial. Not only did there
need to be snow on the ground, but there also needed to be patchy snow (which means
significant freeze/thaw cycles needed to have occurred), and the temperature was
required to be below 0° Celsius (C) due to the inability of passive microwave sensors to
detect wet snow (Walker and Goodison, 1993). The sampling period was minimized to
reduce the effects from any impacts of sublimation or new snowfall on the snow water
equivalent levels of the snowpacks.
To ensure that the three criteria for snow sampling were met, the field campaigns took
place from February 21 st to 23rd, 2005 and on March 2nd, 2008.
4.2
Data Collection
Field observations at each sampling site followed the methods described by Sokol et al.
(1999). In that study, in-situ snow samples were collected along a single transect across
the study area to match the linear path followed by the airborne remote sensing data. The
above approach was modified for use in this study to accommodate the areal (rather than
linear) coverage pattern of spaceborne remote sensing data that were used.
17
The sampled locations were primarily located in agricultural fields along major highways
and grid roads. Snow samples were collected from a variety of land covers that included
fallow and stubble fields, slough margins, shelter belts and pastures. The land cover of
the sampling sites included pastures (Figure 4.1(a)) and shelter belts (Figure 4.1(b)), as
well as fallow (Figure 4.1(c)) and stubble fields (Figure 4.1(d)).
Figure 4.1(a): Pasture
(original in colour)
Figure 4.1(b): Shelter belt
(original in colour)
. «.
u &>*.
4*t-U
M
Figure 4.1(c): Fallow field
(original in colour)
Figure 4.1(d): Stubble field
(original in colour)
18
4.2.1 In-situ Data Collection
A mixture of systematic and random sampling strategies was used to collect the field
data. The study area was systematically divided into 25 km square grid cells, with the
centre of each cell located at 5 km intervals (Figure 4.2). Field measurements were made
as near to the centre of each grid cell as possible. The field campaign was concentrated
into a three-day period to minimize changes in snow pack conditions due to melt or fresh
snowfall.
rh
iri (
ss Wl
Pi3 el
Figure 4.2: Systematic sampling strategy
Two teams, of three to four surveyors each, collected a total of 88 ground observations
from 84 sampling sites that covered an area of 1600 km . As a control, four of the sites
were sampled on consecutive days to ensure data consistency. Included in the 84
sampling sites were 20 sites coincident with an established MSC validation/calibration
snow course data archive (Flight Line 603) (Goodison and Walker, 1994).
The ground-truth data collected from each sampling site included geographic locations,
snow pack profiles, depth measurements, core samples, air and ground temperatures, site
19
photographs, sampling dates and times, and land cover and weather observations.
Sampling site locations were recorded using global positioning system (GPS) handsets.
Since data for differential corrections were not available, the positions have an
approximate accuracy of 10 m.
One of the first observations made at each site was a visual assessment of the percentage
of snow cover. Each team member made this assessment independently and these values
were then averaged to reduce bias in this variable.
For sites with complete snow cover, a total of four snow core samples were collected.
This number was reduced proportionally for partially covered sites. For example, only
two cores were collected from sites determined as having 50% snow coverage, and just
one core was collected from sites with 25% snow cover. This rationale was used to
satisfy the requirement that the cores be taken randomly within each site. Thus, if a site
was found to have 50% snow cover, then the probability of randomly selecting a
sampling location containing snow is only 50%. In this situation, the two core samples
that were not actually collected were simply given zero values for their core lengths,
depths, weights, and densities.
The core samples were obtained using Eastern Snow Conference (ESC-30) snow core
tubes. Measurements from the core samples include the actual depths of the snow packs
from where the cores were removed, the lengths of the cores, and their weights. The
lengths and weights of the cores were used to calculate the ground SWE measurements
(in mm), and the core densities, represented as grams per cubic centimetre (g/cm3). The
20
SWE was calculated by dividing the core weight (in grams) by the cross sectional area of
the ESC-30 snow tube (29.225 cm2) multiplied by 10 (to convert to mm). The SWE
calculation formula is defined as:
SWE (mm) = core weight (g) / 29.225 cm2 * 10
(4.1)
After the SWE measurements were calculated, the core densities were calculated by
dividing the SWE by 10 and also dividing the length of the snow core (to convert to
g/cm3). The formula for the density calculation is defined as:
Density (g/cm3) = SWE /10 / core length
(4.2)
The SWE and density values for the sampling sites were based on the average of the zero
to four core samples collected at each site.
A snow pit was dug at each sampling site and detailed snow pack profiles were made that
included the snow packs' total depths, the number of layers and ice lenses within the
snow packs, the depth and snow grain size of each layer (using Sears snow crystal
screens, containing 1 -3 mm grid patterns), a qualitative description of each layer, and the
air and snow/ground interface temperatures. A total of 16 depth measurements were
made around each snow pit using 15 m long ropes as guides for the purpose of
consistently collecting the depth measurements from 30 m diameter circles. Depth
measurements to the nearest one-half cm were made using 1 m long depth probes. The
depth measurements from each site were used to calculate the average depths within the
sites, which were then used as representative values for the sampling sites. The average
depths are based on the 16 random depth measurements recorded from the circle around
21
the snow pit along with the zero to four depth measurements recorded from the snow core
samples. The SWE for each depth measurement was calculated by multiplying the depth
recording by the average density of the snowpack from the zero to four snow cores at
each sampling site.
Table 3: Measured snow properties used in linear regression models
Independent variable
Variable Type
Description
Percent Snow Cover
Ratio
Depth
Density
Air_Temp
GroundTemp
Ordinal
Ratio
Interval
Interval
Num Layers
Ordinal
Land Cover
IceLens
Nominal
Binary/Nominal
Num_Lenses
Ordinal
Total Lens Thickness
Ordinal
Percentage of snow cover found at each
sampling site.
Mean snow pack depth of each site.
Mean snow pack density of each site.
Air temperature recorded at each site.
Snow/ground interface temperature
recorded from the snow pit.
Number of snow pack layers found in the
snow pit.
Type of land cover.
Indicates whether or not one or more ice
lenses were found in the snow pack of the
snow pit.
Number of ice lenses found in the snow
pack of the snow pit.
Total ice lens thickness found within the
snow pack of the snow pit.
4.3
Remote Sensing Data
Coincident remote sensing data obtained for this study included optical imagery and
passive microwave brightness temperatures from spaceborne sensors with differing
spatial resolutions. Table 4 lists the optical images, which were acquired from the
Landsat 5 Thematic Mapper (TM) sensor, which has a 30 m spatial resolution, and
Terra's MODIS sensor, which has a 500 m spatial resolution.
22
Table 4: Optical imagery used within this study
Data Set
Date(s)
Spatial
Description
Resolution
1) Landsat 5 TM
2) MODIS Snow Cover
Feb 22/05
Mar 2/08
Feb 21/05
Feb 22/05
Feb 23/05
Mar 2/08
30 m Generic Earth scan showing snow
cover
500 m Daily - Fractional Snow Cover
The Landsat data were used for the purposes of identifying the percentage of snow cover,
and classifying the imagery into snow/no snow categories, over a passive microwave
footprint. Two Landsat 5 TM images that coincided with the field campaign dates and
sampling locations were acquired from the United States Geological Survey (USGS)
EarthExplorer web portal.
The MODIS data were obtained as preprocessed snow cover products from the U.S.
National Snow and Ice Data Center (NSIDC). The snow cover data were based on a snow
mapping algorithm that employs a Normalized Difference Snow Index (NDSI) and other
criteria, including atmospheric data (Hall et al., 2006). These data were used to
supplement the fractional snow cover statistics from the Landsat imagery and to quantify
the change of a snow cover that was required to influence a passive microwave derived
SWE estimate.
The passive microwave brightness temperatures were acquired from the U.S. Defense
Meteorological Satellite Program's (DMSP) Special Sensor Microwave/Imager (SSM/I),
which has a 25 km spatial resolution and Aqua's Advanced Microwave Scanning
23
Radiometer - for the Earth Observing System (AMSR-E), which has a 12.5 km spatial
resolution. Table 5 provides a listing of the passive microwave data used in this study.
Table 5: Passive microwave data used within this study
Data Set
Date
Spatial
Description
Resolution
1) 12.5k_AMSR-ESWE
12.5k_AMSR-E 187v
12.5k_AMSR-E 36.5v
2) 25k_AMSR-E SWE
25k_AMSR-E 187v
25k_AMSR-E 36.5v
3) Swath_AMSR-E SWE
Swath_AMSR-E 187v
Swath_AMSR-E 365v
4) 25k_SSM/I SWE
Feb 21/05
Feb 22/05
Feb 23/05
Mar 2/08
Feb 21/05
Feb 22/05
Feb 23/05
Feb 21/05
Feb 22/05
Feb 23/05
Feb 21/05
Feb 22/05
Feb 23/05
Feb 21/05
Feb 22/05
Feb 23/05
Feb 21/05
Feb 22/05
Feb 23/05
Feb 21/05
Feb 22/05
Feb 23/05
Feb 21/05
Feb 22/05
Feb 23/05
Feb 21/05
Feb 22/05
Feb 23/05
Feb 21/05
Feb 22/05
Feb 23/05
12.5 km
SWE estimates derived from AMSRE TB re-sampled to 12.5 km EASEGrid
12.5 km
18.7v TB obtained from AMSR-E resampled to 12.5 km EASE-Grid
12.5 km
36.5v TB obtained from AMSR-E resampled to 12.5 km EASE-Grid
25 km
SWE estimates derived from AMSRE TB re-sampled to 25 km EASEGrid
18.7v TB obtained from AMSR-E resampled to 25 km EASE-Grid
25 km
25 km
36.5v TB obtained from AMSR-E resampled to 25 km EASE-Grid
24 x 12 km
SWE estimates derived from AMSRE TB that have not been re-sampled
24 km
18.7v TB obtained from AMSR-E
that have not been re-sampled
12km
36.5v TB obtained from AMSR-E
that have not been re-sampled
25 km
SWE estimates derived from SSM/I
TB re-sampled to 25 km EASE-Grid
Three of the passive microwave data sets were derived from the brightness temperatures
collected from the AMSR-E sensor. The first AMSR-E data set includes TB re-sampled to
the 12.5 km Equal Area Scalable Earth Grid (EASE-Grid) (Armstrong and Brodzik,
24
1995). The second includes AMSR-E TB re-sampled to the 25 km EASE-Grid, and the
third AMSR-E data set includes non-gridded TB swath data. For comparison, a fourth
data set of SWE estimates derived from Special Sensor Microwave/Imager (SSM/I)
brightness temperatures re-sampled to the 25 km EASE-Grid was included.
4.4
Data Processing and Analysis
4.4.1
In-situ Data Processing
Two data sets, labelled "Snow-Only" and "Actual-Conditions," were created from the
ground data in order to better understand how snow properties over a partial snow cover
are manifested in the remotely sensed SWE estimates. The Snow-Only data set contained
only those observations where there was a measureable snow cover, as determined by
snow depth measurements that were greater than zero. The Actual-Conditions data set
contained data from all site observations, regardless of snow cover. For example, if a site
was found to have snow depth readings of 3, 4, 0, 1 and 4 cm, then the average snow
depth for that site in the Snow-Only data set was recorded as 3.0 cm ((3 + 4 + 1 + 4) / 4),
excluding the zero value. Conversely, the same site in the "Actual-Conditions" data set
would have a mean snow depth of 2.4 cm ((3 + 4 + 0 + 1 + 4 / 5)).
From each of these data sets, four SWE estimates were calculated:
1. "CoreSWE," represented the SWE calculated by using the mean SWE from the
snow cores only;
2. "DerivedSWE," was determined as the average of the Core_SWE and the SWE
derived from the 16 depth measurements;
25
3. "FractionalCoreSWE" was the CoreSWE weighted by the Snow_Cover_
Percent; and
4. "FractionalDerivedSWE"
was
the
DerivedSWE
weighted
by
the
SnowCoverPercent.
These four SWE values were calculated for both the Snow-Only and Actual-Conditions
data sets to produce a suite of eight measures useful in the evaluation of how SWE is
represented over a patchy snow cover.
The ground-truth data were assigned to their corresponding grid locations in a
Geographic Information System (GIS).1 The depth, density, and ground SWE
measurements were represented by the averages of all measurements recorded from each
sampling site. For example, the depth measurement associated with each site is the
average of the 16 random depth measurements recorded from the circle around the snow
pit along with the zero to four depth measurements recorded from the snow core samples.
The ground SWE measurement was the average of the zero to four SWE retrievals from
within the circle. The density was the average SWE divided by the average snow depth.
4.4.2
Remote Sensing Data Processing
All remote sensing images were re-projected to a Universal Transverse Mercator (UTM)
grid to coincide with the ground-truth and base map GIS data. The Landsat imagery were
georeferenced and classified into a Snow/No Snow classification using an inverted
1
ArcGIS V9.2 running under Windows XP SP2
26
Normalized Difference Snow Index (iNDSI). The standard NDSI algorithm was found to
produce images where the snow covered areas appear in darker tones and areas without
snow cover as lighter tones. In the iNDSI, the input band sequence was reversed so that
the snow covered areas were output as white pixels and the non-snow covered areas were
output as black pixels, a more visually intuitive representation. This change was made
simply for visual purposes and had no effect on the results. For Landsat TM imagery,
then, the iNDSI was defined as:
iNDSI = (TM 5 - TM 2) / (TM 5 + TM 2)
(4.3)
The iNDSI data were then manually thresholded into snow-free and snow-covered areas
through interactive manipulation of the displayed image. The accepted NDSI threshold
for snow covered areas is about 0.4 (Rees, 2006). This value is known to be heavily
dependent on seasonality, image quality, and cloud cover.
The MODIS data were re-projected from their native Sinusoidal projection into a UTM
North American Datum (NAD) 1983 (or UTM NAD 83) coordinate system and imported
into the GIS to be compared with the other data sets. As the MODIS images were already
available as fractional snow cover data, there was no need to further process them.
The passive microwave images were obtained as raster arrays with grid spacings varying
from 12.5 km to 25 km (Table 5). In order to combine these data with the optical imagery
of considerably higher spatial resolutions (30 m and 500 m), each passive microwave
image pixel centroid was considered as a point (Figure 4.3). Passive microwave imagery
pixel footprints were then simulated in the GIS by creating Thiessen polygons (Figure
27
4.4). The Thiessen polygon algorithm segments the measurement space such that every
polygon encloses the region closest to its pixel centroid (O'Sullivan and Unwin, 2003).
Although the algorithm does not produce perfectly squared pixels, the fact that the
algorithm measures the mid-points between the pixel centroids ensures that the spatial
resolutions of the remote sensing data sets are preserved. Passive microwave footprints
were created at 12.5 km spacing for the AMSR-E data and at 25 km spacing for both the
AMSR-E and SSM/I imagery.
Figure 4.3: Pixel centroids
Figure 4.4: Theissen polygon footprints
SWE values representative of each cell were calculated for the AMSR-E data using
MSC's operational prairie algorithm (Formula 3.2). The SSM/I data did not require
further processing since they were obtained from the MSC as prairie algorithm SWE
values.
Finally, the optical images were used to calculate the percentage of snow cover across the
passive microwave footprints. All Thiessen polygons representing the passive microwave
28
footprints were transferred from the GIS to a remote sensing image analysis system
where they were aligned with the TM imagery and the percent of snow cover for each
footprint was determined. This percentage was then used to calculate the weighted SWE
for each pixel. Formula 5.1 shows the weighted SWE algorithm, where w represents the
satellite-derived percentage of snow cover for each pixel.
SWEW (mm) = (-20.7 - (37v - 19v) * 2.59) w
4.4.3
(4.4)
Map Queries
Remotely sensed SWE estimates were compared in the GIS to ground SWE
measurements that had been averaged over the pixel footprints. Differences within +/- 20
mm of ground measurements were considered to be equivalent - since this was the
known accuracy of the MSC SWE algorithm (Goodison and Walker, 1994; Derksen et
ah, 2003a; 2003b). SWE estimates found not to be within the +/- 20 mm threshold were
considered as over- or underestimated.
4.5
Statistical Tests
Z-tests were performed between the results of the Snow-Only and Actual-Conditions data
sets to determine if there were significant differences in algorithm performance between
the two ground-truth representations. Linear regression models were developed between
2
ENVI 4.2 running under Windows XP SP2
29
the remote sensing data (as the dependent variables) and the ground observations from
the coincident sampling sites (as the independent variables).
The linear regression outputs were interpreted following the systematic procedure
proposed by Gupta (2000) (Figure 4.5). The significance of the model fit was then
analyzed. The model significance explains the deviations of the dependent variables (e.g.,
the SWE estimates and TB). A model significance of 0.10 (90% confidence level) was
used as the cutoff for model acceptance. Models with levels below the 90% confidence
level were removed from further analysis.
Input Remote Sensing data as
dependent variable and
all in-situ data as
Independent variables
^ ^ - - ^
Check significance of model fit
If significance < 0.10, then model fits
the data at 90% level of confidence
If significance » .010, then throw out
the model
Analyze Adjusted R*
explains proportion of the
variance in remote sensing
data that are explained by the
variations of the in-situ
measurements
Identify reliability of individual coefficients
for independent variables
If significance < 0,10, then in-situ variable is
significant at 90% level of confidence
If significance > .010, then in-situ variable
is irrelevant
Figure 4.5: Regression analyses flow chart
Linear regressions were performed using the Statistical Package for the Social Sciences (SPSS) software.
30
The next step in interpreting the regression output was to analyze the Adjusted R value
from the model summary. This value is sensitive to the addition of irrelevant variables,
and is a measure of the proportion of the variance in the dependent variables that are
explained by the variations of the independent variables. For example, an Adjusted R
value of 0.50 suggests that 50% of the variance in a SWE estimate is explained by the
variation in the ground-truth measurements.
The third, and final, interpretation step involved identifying the reliability of the
individual coefficients for the independent variables. The Beta values included in the
coefficients output indicate the predicted coefficients for the model, along with their
standard errors and significances. Similar to the significance of the model fit, if a
coefficient results in a significance value above 0.10, then it can be concluded that the
independent variable is not significant at a 90% level of confidence.
31
CHAPTER FIVE - RESULTS AND DISCUSSIONS
In the first part of this chapter, the results from the statistical comparisons between the
ground data sets and the coincident passive microwave data are presented. The second
part includes the results of the spatial comparisons of the optical imagery and the passive
microwave data.
5.1
Statistical Analyses
5.1.1
In-situ vs. Passive Microwave SWE Estimates
When the passive microwave SWE estimates were compared with the ground
measurements, the SSM/I and swath AMSR-E SWE data provided the closest
approximations. This was expected, because: i) the MSC algorithm used to derive SWE
estimates was actually developed for SSM/I TB; and ii) the swath AMSR-E TB had not
been re-sampled, thus, they were truer representations of the interaction between the
sensor and the ground surface. Tables 6 and 7 illustrate the number (and percentage in
brackets) of sampling sites that were found to be equivalent (i.e., within +/- 20 mm
SWE) to the remotely sensed SWE estimates for the Snow-Only and Actual-Conditions
data sets. These results also show a slight increase in algorithm agreement for the SnowOnly data set (i.e., when only the amount of snow found at each sampling site was
included in the ground-truth observations).
Table 6: Number of equivalent Snow-Only coincident sites (n=88)
SWE Calculation
Core SWE
Fractional Core SWE
Derived SWE
Fractional Derived SWE
SSM/I
55 (63%)
50 (57%)
58 (66%)
45(51%)
12.5k AMSR-E
42 (48%)
29 (33%)
32 (36%)
24 (27%)
32
25k AMSR-E
20 (23%)
11(13%)
15 (17%)
7 (8%)
Swath AMSR-E
56 (64%)
49 (56%)
54(61%)
45 (51%)
Table 7: Number of equivalent Actual-Conditions coincident sites (n=88)
SWE Calculation
Core SWE
Fractional Core SWE
Derived SWE
Fractional Derived SWE
SSM/I
51 (58%)
44 (50%)
45(51%)
33 (38%)
12.5k AMSR-E
34 (39%)
27(31%)
26 (30%)
23 (26%)
25k AMSR-E
13(15%)
10(11%)
9(10%)
7 (8%)
Swath AMSR-E
52 (59%)
45(51%)
49 (56%)
40 (45%)
By weighting the in-situ SWE values by the percentage of snow cover found at the sites
(i.e. reading down each column) the agreement with the remote sensing estimates
decreased by an average of approximately 10% in the Snow-Only data set, and by an
average of approximately 7% in the Actual-Conditions data set. Further, the remote
sensing SWE algorithm generally had a higher level of agreement with CoreSWE
measurements than with DerivedSWE values. Therefore, for a patchy snow cover, the
MSC SWE algorithm appeared to have the closest agreement with ground SWE
measurements based only on the core samples
Further analyses using only the CoreSWE measurements found that, on average, the
remote sensing algorithm tended to overestimate the patchy in-situ SWE measurements
in all cases (Table 8). This was not surprising since the remote sensing algorithm was
originally derived for a complete snow cover.
Table 8: Mean differences in SWE between remote sensing estimates and in-situ
measurements (for core samples only - all units are in mm)
Snow-Only core_swE
Actual-Conditions core SWE
SSM/I
4.8
8.7
12.5k AMSR-E
17.7
21.6
25k AMSR-E
32.1
35.9
Swath AMSR-E
5.4
9.3
When the effects that varying land covers had on the spaceborne SWE estimates were
examined, the ranges in ground SWE measurements were very high. This was
particularly true for sampling sites that included shelter belts and fallow fields, which
33
were found to have drastically different snow conditions than stubble fields and pastures.
Table 9 shows that there was little difference in the mean SWE values representing
stubble fields (26.1 mm) and pastures (23.0 mm), but great disparity between these values
and fallow fields (1.7 mm) and shelter belts (94.8 mm).
Table 9: Actual-Conditions SWE, depth and density measurements by land cover
Land Cover
Stubble
n
54
Fallow
15
Pasture
12
Shelter Belt
5.1.2
2
SWE (mm)
mean = 26.1
min =1.3
max = 66.3
mean = 1.7
min = 0
max = 8.8
mean = 23.0
min = 5.1
max = 44.1
mean = 94.8
min = 75.6
max= 114.0
Depth (cm)
mean = 7.9
min = 1.0
max =19.0
mean = 0.8
min = 0
max = 4.7
mean = 6.7
min= 1.8
max= 16.4
mean = 26.8
min = 22.7
max = 30.9
Density (g/cm3)
mean = 0.226
min= 0.026
max = 0.511
mean = 0.054
min = 0
max = 0.265
mean = 0.219
min = 0.110
max = 0.364
mean = 0.292
min = 0.267
max = 0.316
Linear Regression Models
Linear regressions were calculated using the remotely sensed SWE estimates and
brightness temperatures as the dependent variables and the in-situ observations as the
independent variables. The regressions were determined separately for the Snow-Only
and Actual-Conditions data sets.
Table 10 provides an instructive example of how an output coefficient table was
analyzed. The SWE predicted by the model was specified by the Constant's Beta
coefficient, 30.5 mm. The standard error of this prediction was 1.9 mm of SWE. The next
step was to identify the significance of the independent variables' Beta coefficients. In
this example, a coefficient of .149 for the average snow pack depth was found to be
34
significant at a 95% level of confidence (significance = .043), but the coefficient of 2.899 for the average density was irrelevant (significance = .776) towards the predicted
SWE value. Therefore, in this example, the predicted SWE was significantly associated
with depth, but not with density.
Table 10: Example coefficient table output
Model
Constant (predicted SWE value)
Depth
Density
Unstandardized
Beta
30.5
.149
-2.899
Coefficients
St. Error
1.9
.111
10.148
SIGNIFICANCE
—
.043
.776
Table 11 shows the regression results between the Swath AMSR-E SWE estimates and
both the Snow-Only and Actual-Conditions data sets. The Model Fit shows that both data
sets match the satellite estimates at 99% levels of confidence. From the Unstandardized
Beta Coefficients it can be seen that the only significant variables are the percentage of
snow cover, and whether or not one or more ice lenses were found in the snow pit.
Interestingly, density in the Snow-Only data set appears not to make a significant
contribution, while it is found to be significant at a 90% confidence level in the ActualConditions data set. The Model Summaries indicate that the proportion of the variance in
the satellite estimates that is explained by the ground observations are just 15.8% and
17.9% for the Snow-Only and Actual-Conditions data sets, respectively.
35
Table 11: Regression results between Swath AMSR-E SWE estimates and in-situ
observations (significant coefficients are shown in bold italics)
Snow-Only
Swath_AMSR-E SWE
1) Model Fit
2
2) Model Summary (Adjusted R )
3) Coefficients
Actual-Conditions
.008
.004
.158
.179
Beta
St. Error
Sig.
Beta
St. Error
Sig.
32.646
2.280
—
32.552
2.140
—
8.508
4.276
.050
10.079
4.310
.022
.143
.220
.518
.187
.216
.390
-6.009
6.646
.369
-19.131
11.357
.096
•
Constant(Predicted SWE)
•
Snow_Cover_Percent
•
Depth
•
Density
•
Air_Temp
-.050
.306
.870
-.026
.302
.932
•
GroundTemp
-.162
.492
.743
-.257
.489
.601
•
Num_Layers
.371
1.500
.805
.691
1.498
.646
•
LandCover
-1.555
1.376
.262
-1.490
1.355
.275
•
Ice_Lens
-7.354
3.369
.032
-6.738
3.267
.043
•
Num Lenses
.563
3.205
.861
-.151
3.204
.963
•
Total_Lens_Thickness
.108
.647
.868
.214
.643
.740
Table 11 also shows that snow pack densities from the Actual-Conditions data set were
significantly positively correlated with SWE estimates derived from the Swath AMSR-E
data. This was expected since dense and complex snow packs have been shown to
amplify scattering and tend to produce remotely sensed SWE overestimates (Sokol et al.,
1999). The Swath AMSR-E imagery was the only remote sensing data set to show such a
statistically significant correlation, likely because — since they had not been re-sampled
to the EASE-Grid — these data were closer representations of the Earth radiances
originally detected by the sensor.
Similar regressions were run between all of the remote sensing and in-situ data sets. The
results are summarized in Table 12. Regression models derived for the 12.5 km AMSR-E
is.7v and 25 km AMSR-E swEdata were not statistically significant. Regressions from the
36
AMSR-E TB found that the 18.7v TB resulted in having more significant variables than
those collected from the 36.5v TB. While the Snow_Cover_Percent, and IceLens
variables were found to be significant in both TB regressions, Depth and GroundTemp
were found to also be significant in the 18.7v TB regression. With the 12.5 km AMSR-E
data it is interesting to note that in comparison to the non-gridded Swath_AMSR-E
analyses a completely different set of variables, except for the binary variable IceLens,
was found to be significant. The significant variables in this data set include: Depth,
AirTemp, LandCover, and IceLens. However, as these data had been re-sampled,
there was less confidence in these regression results compared to those of the swath
results. Unlike the 12.5 km AMSR-E regression results, the 25 km AMSR-E ig.7V TB were
found to be significant, and the 36.5v TB were marginally significant. The 25 km SSM/I
SWE regressions produced nearly identical results between the Snow-Only and ActualConditions data sets.
37
SO: Snow-Only
AC: Actual Conditions
1) Model Fit
I) Model Summary
(Adjusted R 2 )
5) Coefficients
• Snow_Cover_Percent
• Depth
' Density
1
Air Temp
• GroundTemp
• NumLayers
• LandCover
• Ice Lens
• Num_Lenses
• Total Lens Thickness
0.2
•
0.2
Swath
AMSR-E
SWE
SO AC
•
•
•
•
0.3
0.3
Swath
AMSR-E
18.7v
SO AC
0.2
•
0.3
Swath
AMSR-E
36.5v
SO AC
•
•
•
0.3
•
•
•
0.3
12.5k
AMSR-E
SWE
SO AC
12.5k
AMSR-E
18.7v
SO AC
•
•
•
•
•
•
•
•
12.5k
AMSR-E
36.5v
SO AC
•
•
0.3 0.3
25k
AMSR-E
SWE
SO AC
Table 12: Regression results between remotely sensed SWE estimates and in-situ observations
(• denotes a statistically significant correlation)
0.3
0.3
25k
AMSR-E
18.7v
SO AC
•
•
0.1
•
•
0.1
25k
AMSR-E
36.5v
SO AC
0.1
•
0.1
•
25k
SSM/I
SWE
SO AC
5.2
Spatial Analysis Results
5.2.1
Fractional Snow Cover from Thematic Mapper Imagery
The Landsat Thematic Mapper (TM) Snow/No Snow binary classification images
produced comparable results to the colour composite images (Figures 5.1 and 5.2). An
iNDSI value of 0.22 was used to threshold the classified image into snow and no-snow
categories. The 0.22 iNDSI value was determined inductively by using various threshold
values beginning with 0.4, which had previously been found to be representative of snow
cover (Rees, 2006). Various threshold values were attempted until a visual match of snow
cover was found between the Standard False-Colour image and the iNDSI image. There
was a strong agreement between the pixels classified as snow-covered in the TM image
and the snow cover observed in the field.
Figure 5.1: Landsat TM 2005 Standard FalseColour Composite (TM Bands 4, 3, 2, shown as
red, green, and blue, respectively)
(original in colour)
39
Figure 5.2: Landsat TM 2005 iNDSI Snow
Classification (Snow-covered areas shown in
white) (Snow/No Snow)
To verify the results from the 2005 field data, a second field campaign was undertaken in
2008. Figures 5.3 and 5.4 show the results from the 2008 TM image processing using the
same procedures used on the 2005 TM imagery. The iNDSI threshold value used to
classify the 2008 image was 0.20, which is close to the 2005 value. The 2008 iNDSI
value was arrived at inductively, using the same procedure used to determine the 2005
iNDSI value. As in 2005, the 2008 field observations visually correlated well with the
classified imagery. Therefore, it would appear that a value of approximately 0.20 is an
acceptable threshold for the southern Saskatchewan region in mid-late winter using TM
bands 2 and 5 for an inverted NDSI algorithm.
Figure 5.3: Landsat TM 2008 Standard FalseColour Composite (TM Bands 4, 3, 2, shown as
red, green, and blue, respectively)
(original in colour)
Figure 5.4: Landsat TM 2008 iNDSI Snow
Classification (Snow-covered areas shown in
white) (Snow/No Snow)
The Snow/No Snow classified TM imagery was integrated with coincident AMSR-E and
SSM/I footprints in the GIS to determine the impact of fractional snow cover on passive
40
microwave SWE retrievals. Figure 5.5 shows the AMSR-E and SSM/I footprints
(represented as Thiessen polygons) overlaying the TM image of the study area. The
footprints highlighted with the white boundaries represent the six 12.5 km AMSR-E
footprints used in this study, while the footprint with the black boundary represents both
of the AMSR-E and SSM/I 25 km footprints.
I
I 12.5 km AMSR-E
Figure 5.5: AMSR-E and SSM/I pixels on Landsat TM 2005 Standard
False-Colour Composite background
(original in colour)
41
5.2.2
25 km SSM/I SWE Weighted by TM Snow Fractions
Figure 5.6 shows the unweighted (left) and weighted (right) 25 km SSM/I pixel from the
afternoon orbit of February 21, 2005. The fractional snow cover analysis found this
footprint had a 59% snow cover. Ground-truth observations collected from forty-two
well-distributed and spatially representative sampling sites within the pixel yielded an
average in-situ SWE measurement of 27.3 mm. Of the forty-two sampling sites, twentyfive (60%) were found to be within the +/- 20 mm accuracy of the MSC SWE algorithm
(Goodison and Walker, 1994; Derksen et al., 2003a; 2003b). The sites found to be within
the accuracy threshold are symbolized as solid green circles in the images and labelled
"Equivalent Sites," while the sites found to be outside the threshold are labelled "NonEquivalent Sites" and symbolized as solid red circles. The unweighted SSM/I SWE
estimate was 27.7 mm, which was nearly identical to the estimated +/- 20 mm accuracy
of the MSC SWE algorithm.
The weighted SSM/I estimate was 16.3 mm, a value which was further from the ground
measurement for the SSM/I orbit date shown in the image. However, when the snow
cover percent weight was applied to the SSM/I SWE estimates, thirty-one of the fortytwo (74%) sampling sites were found to be within the weighted SSM/I estimates, which
was a 14% improvement. The image on the right in Figure 5.6 illustrates the
improvement of the weighted algorithm by showing a higher number of equivalent sites
(green circles) than the unweighted image shown on the left.
42
Figure 5.6: SSM/125 km TM Fractional Snow Cover for February 21-23, 2005
(unweighted on left - weighted on right)
(original in colour)
The SSM/I data from February 21 through February 23, 2005 are presented in Figure 5.7.
The dashed green line in Figure 5.7, representing the weighted SWE estimates, follows
the ground SWE (bold black line) more closely than the unweighted SWE estimates (red
line) for satellite data from morning orbits. In this example, the weighted SWE algorithm
improved the accuracy of the SWE estimates on a patchy snow covered area for two of
the five coincident SSM/I imagery dates, specifically, the morning orbits.
43
Figure 5.7: SSM/125 km SWE vs. TM-Weighted SWE graph for February 21-23,2005
(original in colour)
5.2.3
25 km AMSR-E SWE Weighted by TM Snow Fractions
Figure 5.8 shows the unweighted (left) and weighted (right) 25 km AMSR-E pixel from
the orbit of February 22, 2005. As the Thiessen polygon represented the same pixel as the
25 km SSM/I shown in Figure 5.6, the attributes of the optical imagery and the sampling
sites were also the same. Unlike the 25 km SSM/I pixel, the unweighted AMSR-E SWE
estimate was overestimated, at 58.3 mm, significantly outside the range of the estimated
+/- 20 mm accuracy of the MSC SWE algorithm. Of the forty-two sampling sites, only
ten (24%) were found to be equivalent to the AMSR-E estimate.
The weighted SWE estimate was 34.4 mm, a value that lowered the SWE estimate to
within the acceptable threshold of the ground measurement. When the snow cover
percent weight was applied to the AMSR-E SWE estimates, twenty-seven of the forty44
two (64%) sampling sites were found to be within the weighted AMSR-E estimates,
which was a 40% improvement. The image on the right in Figure 5.8 illustrates the
significant improvement of the weighted algorithm by showing a much higher number of
equivalent sites than the unweighted image shown on the left.
Figure 5.8: AMSR-E 25 km TM Fractional Snow Cover for February 21-23, 2005
(unweighted on left - weighted on right)
(original in colour)
The 25 km AMSR-E data from February 21 through February 23, 2005 are shown in
Figure 5.9. The dashed green line in Figure 5.9, representing the weighted SWE
estimates, follows the ground SWE (bold black line) more closely than the unweighted
SWE estimates (red line). In this example, the weighted SWE algorithm improved the
accuracy of the SWE estimates on a patchy snow covered area for all three of the
45
coincident AMSR-E imagery dates. This improvement is considered significant, as the
unweighted SWE was overestimated from the ground measurements by up to 34 mm.
AMSR-E (25 km) Pixel 1, February 21-23, 2005
SWE Estimate vs. Weighted SWE Estimate
70.0 i
^ ^ - • S W E Estimate (mm)
- » - Weighted SWE Estimate (mm)
" • • • M e a n Ground SWE (mm)
^ • ™ » U p p e r 20 mm Threshold
^ ^ — • L o w e r 20 mm Threshold
60.0
_
50.0 -
E
+ 24) mm Threshold
g
40
1
>
-° "
"5
oUJ
S
30.0 -
CO
g
§
=
20.0
- 20 mm Threshold
10.0 •
0.0
Date (Time)
Figure 5.9: AMSR-E 25 km SWE vs. TM-Weighted SWE graph for February 21-23,2005
(original in colour)
5.2.4
12.5 km AMSR-E SWE Weighted by TM Snow Fractions
There were six AMSR-E satellite image pixels at 12.5 km spatial resolution that covered
the study area. As shown in Figure 5.5, these were labeled as Pixel 1 through Pixel 6.
The image in Figure 5.10 shows the unweighted and weighted SWE images for Pixel 1,
the first of the six 12.5 km AMSR-E pixels. The TM-derived fractional snow cover of
this footprint was 45%. Ground-truth observations collected from eleven sampling sites
within the pixel yielded an average in-situ SWE measurement of 33.8 mm. Of the eleven
sites, nine (82%) were found to be equivalent to the AMSR-E SWE estimate. The
46
unweighted SSM/I SWE estimate was 19.4 mm, which was within the estimated +/- 20
mm accuracy of the MSC SWE algorithm. However, the weighted SWE estimate was 8.7
mm, a value that was further from the ground measurement. When the snow cover weight
was applied to AMSR-E estimate, only seven of the eleven (64%) sampling sites were
found to be within the algorithm accuracy. This discrepancy is likely the result of the
unevenly distributed and spatially unrepresentative distribution of ground measurements
from this day (Figure 5.10). In this instance, it was impossible to accurately validate the
satellite-derived SWE with in-situ data. Results from the unweighted algorithm are
shown on the left in Figure 5.10, while the weighted algorithm results are shown on the
right.
Figure 5.10: AMSR-E 12.5 km Pixel 1 TM Fractional Snow Cover for February 22, 2005
(unweighted on left - weighted on right)
(original in colour)
47
Figure 5.11 presents the AMSR-E 12.5 km Pixel 1 data from February 21 through
February 23, 2005. The AMSR-E SWE estimates for this pixel more closely matched the
ground measurements than the weighted SWE estimates. Since the ground measurements
were not spatially representative, however, the weighted estimates may have been closer
to the actual values.
AMSR-E (12.5 km) Pixel 1, February 21-23, 2005
SWE Estimate vs. Weighted SWE Estimate
70.0
• S W E Estimate (mm)
-Weighted SWE Estimate (mm)
• Mean Ground SWE (mm)
• Upper 20 mm Threshold
• Lower 20 mm Threshold
+ 20 mm Threshold
30.0
20.0
- 2 0 mm Threshold
10.0
0.0
Date (Time)
Figure 5.11: AMSR-E 12.5 km Pixel 1 SWE vs. TM-Weighted SWE graph
for February 21-23, 2005
(original in colour)
Figure 5.12 shows the second of the unweighted and weighted 12.5 km AMSR-E pixel.
The TM-derived fractional snow cover of this footprint was 58%. Ground-truth
observations collected from nine relatively well-distributed and spatially representative
sampling sites within the pixel yielded an average in-situ SWE measurement of 23.4 mm.
Of the nine sampling sites, only three (33%) were found to be within the MSC SWE
48
algorithm accuracy threshold. The unweighted AMSR-E SWE estimate was 39.5 mm,
which slightly overestimated the +/- 20 mm accuracy.
The weighted AMSR-E estimate was 22.9 mm, a value that was much closer to the
ground measurement. When the snow cover percent weight was applied to the AMSR-E
SWE estimates, five of the nine (56%) sites were found to be within the weighted
AMSR-E estimates, a 23% improvement. The image on the right in Figure 5.12 illustrates
the improvement of the weighted algorithm by showing a higher number of equivalent
sites than the unweighted image on the left.
Figure 5.12: AMSR-E 12.5 km Pixel 2 TM Fractional Snow Cover for February 23,2005
(unweighted on left - weighted on right)
(original in colour)
Figure 5.13 shows the AMSR-E 12.5 km Pixel 2 data from February 21 and February 23,
2005. There were no ground data collected on February 22, 2005 so analysis of this pixel
49
is limited to two dates. The weighted AMSR-E SWE estimates for this pixel more closely
matched the ground measurements than the unweighted SWE estimates. The weighted
SWE algorithm in this example improved the accuracy of the SWE estimates for both of
the coincident AMSR-E imagery dates to a value nearly identical to the ground SWE
measurement and lowered the overestimated SWE from February 21, 2005 to within the
+/- 20 mm accuracy of the MSC SWE algorithm.
Figure 5.13: AMSR-E 12.5 km Pixel 2 SWE vs. TM-Weighted SWE graph
for February 21 and 23, 2005
(original in colour)
Figure 5.14 presents the third 2005 12.5 km unweighted and weighted AMSR-E pixel,
captured from the orbit of February 22, 2005. The TM-derived fractional snow analysis
found this footprint had an 89% snow cover. Ground-truth observations collected from
five relatively well-distributed and spatially representative sampling sites within the pixel
50
yielded an average in-situ SWE measurement of 32.9 mm. Of the five sampling sites,
three (60%) were found to be within the MSC SWE algorithm accuracy of +/- 20 mm.
The unweighted AMSR-E SWE estimate was 54.3 mm, which was again found to be
over the estimated accuracy of the MSC algorithm.
The weighted SWE estimate was 48.3 mm, a value that was slightly closer to the ground
measurement. No improvement was visually evident when the snow cover percent weight
was applied to the AMSR-E SWE estimates, as the same three (60%) sites were found to
be within the weighted AMSR-E estimate. The image on the right in Figure 5.14 shows
the weighted algorithm results, while the unweighted results are on the left.
Figure 5.14: AMSR-E 12.5 km Pixel 3 TM Fractional Snow Cover for February 22, 2005
(unweighted on left - weighted on right)
(original in colour)
51
The data from February 21 through February 23, 2005 are shown in Figure 5.15.
Although a visual analysis of Figure 5.14 did not show an improvement in the weighted
SWE estimates, the graph in Figure 5.15 illustrates that the weighted estimates for this
pixel more closely matched the ground measurements than the unweighted SWE
estimates. The weighted SWE algorithm in this example improved the accuracy of the
SWE estimates for all three of the coincident AMSR-E imagery dates, lowering the SWE
estimates to within the +/- 20 mm accuracy of the MSC SWE algorithm.
A M S R - E (12.5 km) Pixel 3, February 21-23, 2005
S W E Estimate vs. Weighted SWE E s t i m a t e
70.0 -i
^ ^ — S W E Estimate (mm)
—
-Weighted SWE Estimate (mm)
« ^ ^ » M e a n Ground SWE (mm)
60.0 -
• ^ ^ U p p e r 20 mm Threshold
^ ^ ^ L o w e r 20 mm Threshold
_
50.0 -
+ 20 mm Threshold
!
c
•=
>
40.0
"5
3
£
30.0 -
-20 mm Threshold
v> 20.0 -
10.0
0.0 •
Date (Time)
Figure 5.15: AMSR-E 12.5 km Pixel 3 SWE vs. TM-Weighted SWE graph
for February 21-23,2005
(original in colour)
The fourth 2005 12.5 km unweighted and weighted AMSR-E pixel is shown in Figure
5.16. The TM-derived fractional snow cover of this footprint was 52%. Ground-truth
observations collected from ten relatively well-distributed and spatially representative
52
sampling sites within the pixel yielded an average in-situ SWE measurement of 27.9 mm.
Of the ten sampling sites, seven (70%) were found to be within the MSC SWE algorithm
accuracy. The unweighted SWE estimate was 36.0 mm, which was found to be within the
+/- 20 mm accuracy of the MSC algorithm.
The weighted AMSR-E estimate was 18.7 mm, which was found to be closer to the
ground measurement than the unweighted SWE. A slight improvement was visually
apparent when the snow cover percent weight was applied to the AMSR-E SWE
estimates, as eight of the ten (80%) sampling sites were found to be within the weighted
AMSR-E estimate. The image on the right in Figure 5.16 shows improvement in the
weighted algorithm results over the unweighted results on the left.
Figure 5.16: AMSR-E 12.5 km Pixel 4 TM Fractional Snow Cover for February 23, 2005
(unweighted on left - weighted on right)
(original in colour)
53
Figure 5.17 presents the AMSR-E 12.5 km Pixel 4 data from February 21 and February
23, 2005. Similar to Pixel 2, there were no ground data collected within the AMSR-E
footprint on February 22, 2005 so the analysis of this pixel is limited to two dates.
Although the unweighted AMSR-E estimates for this pixel are within the MSC algorithm
threshold, they slightly overestimated the ground SWE measurements. The weighted
SWE estimates are also within the MSC algorithm threshold, but they are approximately
20 mm lower than the unweighted SWE estimates. In this example, the weighted
algorithm did not significantly improve the accuracy of the SWE estimates for either of
the coincident AMSR-E imagery dates, but it did decrease the SWE estimates by 20 mm.
AMSR-E (12.5 km) Pixel 4, February 21-23, 2005
SWE Estimate v s . Weighted S W E Estimate
70.0
• • SWE Estimate (mm)
—
—Weighted SWE Estimate (mm)
^ ^ i " » M e a n Ground SWE (mm)
60.0
• • • " • " U p p e r 20 mm Threshold
• • " ^ " L o w e r 20 mm Threshold
?
50.0
+ 20 flim Threshold
E.
•5
4 0 0
>
3
111
3 30.0
n
s
I
-20 mm Threshold
w 20.0
10.0
0.0
>
Date (Time)
Figure 5.17: AMSR-E 12.5 km Pixel 4 SWE vs. TM-Weighted SWE graph
for February 21 and 23,2005
(original in colour)
54
The fifth 12.5 km unweighted and weighted AMSR-E pixel, shown in Figure 5.18, was
collected from the orbit of February 23, 2005. The TM-derived fractional snow cover
analysis found 39% of this footprint to be covered in snow. Ground-truth observations
collected from four relatively well-distributed and spatially representative sampling sites
within the pixel yielded an average in-situ SWE measurement of 15.3 mm. Of the four
sites, only one (25%) was found to be within the MSC SWE algorithm accuracy of+/- 20
mm. The unweighted SWE estimate was 38.3 mm, which appears to have significantly
overestimated the accuracy of the MSC SWE algorithm.
The weighted estimate was 14.9 mm, which was much closer to the ground measurement,
and decreased the SWE estimate to within the accuracy threshold. A significant visual
improvement was apparent when the snow cover percent weight was applied to the
AMSR-E SWE estimates, as three of the four (75%) sampling sites were found to be
within the weighted AMSR-E estimate. The image on the right in Figure 5.18 shows the
improvement in the weighted algorithm results over the unweighted results presented in
the image on the left.
55
Figure 5.18: AMSR-E 12.5 km Pixel 5 TM Fractional Snow Cover for February 23, 2005
(unweighted on left - weighted on right)
(original in colour)
Figure 5.19 shows the AMSR-E 12.5 km Pixel 5 data from February 23, 2005. Ground
data were not collected on either February 21 or 22, 2005 from within the AMSR-E
footprint. Thus, analysis of this pixel is limited to one date. For visual purposes, the data
for the single date in Figure 5.19 are duplicated in the graph. The weighted SWE estimate
for this pixel more closely matched the ground measurement than the unweighted SWE
estimate. The weighted algorithm in this example significantly improved the accuracy of
the SWE estimate for the coincident AMSR-E imagery date and lowered the
overestimated SWE to within the +/- 20 mm accuracy of the MSC algorithm. However,
caution should be exercised with respect to the analysis of AMSR-E Pixel 5, as ground
data from only four sampling sites were coincident with the footprint.
56
AMSR-E (12.5 km) Pixel 5, February 21-23, 2005
SWE Estimate vs. Weighted SWE Estimate
- S W E Estimate (mm;
•Weighted SWE Estimate (mm)
• Mean Ground SWE (mm)
60.0
• Upper 20 mm Threshold
•Lower 20 mm Threshold
40.0
+ 20 mm Threshold
£ 30.0
0) 20.0
10.0
- 20 mm Threshold
0.0
Date (Time)
Figure 5.19: AMSR-E 12.5 km Pixel 5 SWE vs. TM-Weighted SWE graph
for February 23, 2005
(original in colour)
The sixth 12.5 km unweighted and weighted AMSR-E pixel, is presented in Figure 5.20.
The TM-derived fractional snow analysis found a 67% snow cover within this footprint.
Ground-truth
observations
collected
from
seven well-distributed
and
spatially
representative sampling sites within the pixel yielded an average in-situ SWE
measurement of 16.6 mm. Of the seven sampling sites, five (71%) were found to be
within the accuracy of the MSC algorithm. The unweighted SWE estimate was 32.4 mm,
which is within the accuracy of the MSC SWE algorithm.
The weighted estimate was 21.7 mm, which was closer to the ground measurement. A
slight visual improvement was apparent when the snow cover percent weight was applied
to the AMSR-E SWE estimates, as six of the seven (86%) sampling sites were found to
57
be within the weighted AMSR-E estimate. The image on the right in Figure 5.18 shows
the visual improvement in the weighted algorithm results over the unweighted results
presented in the image on the left.
Figure 5.20: AMSR-E 12.5 km Pixel 6 TM Fractional Snow Cover for February 22, 2005
(unweighted on left - weighted on right)
(original in colour)
Figure 5.21 presents the AMSR-E 12.5 km Pixel 6 data from February 22 and February
23, 2005. In-situ data were not collected from within the AMSR-E footprint on February
21, 2005. The weighted SWE estimates for this pixel more closely matched the ground
measurements than the unweighted SWE estimates. The weighted algorithm in this
example improved the accuracy of the SWE estimates for both of the coincident imagery
dates to a value closer to the ground measurement.
58
AMSR-E (12.5 km) Pixel 6, February 21-23, 2005
S W E Estimate vs. Weighted SWE Estimate
70.0
T
•
* * — * SWE Estimate (mm)
—
-Weighted SWE Estimate (mm)
^ • ™ » M e a n Ground SWE (mm)
60.0
••miii i Upper 20 mm Threshold
in
in Lower 20 mm Threshold
50.0
I
- 26 mm threshold
«*
Date (Time)
Figure 5.21: AMSR-E 12.5 km Pixel 6 SWE vs. TM-Weighted SWE graph
for February 22 and 23,2005
(original in colour)
5.2.5
Fractional Snow Cover from MODIS Imagery
Coincident MODIS images were obtained as preprocessed Fractional Snow Cover
products from the U.S. National Snow and Ice Data Center (NSIDC). The snow cover
data were based on a snow mapping algorithm that employs a Normalized Difference
Snow Index (NDSI), and incorporates data recorded from other bands on the MODIS
sensor for the purpose of masking certain features, such as clouds and waterbodies,
within the finished product (Hall et al., 2006). The MODIS data were used in this study
to supplement the fractional snow cover statistics from the TM imagery and to quantify
the change of a snow cover that was required to influence a passive microwave-derived
SWE estimate.
59
As with the fractional snow cover analyses performed with the TM imagery, the MODIS
images were also integrated with coincident AMSR-E and SSM/I footprints in a GIS to
determine the impact of fractional snow cover on passive microwave SWE retrievals.
Figure 5.22 shows the AMSR-E and SSM/I footprints overlaying the MODIS Fractional
Snow Cover image from the orbit of February 23, 2005. The footprints highlighted with
the light grey boundaries represent the 12.5 km AMSR-E footprints, while the footprint
with the black boundary represents both of the AMSR-E and SSM/I 25 km footprints.
Also shown in Figure 5.22 are classes from the MODIS Fractional Snow Cover product.
To reduce complexity in analyzing the MODIS imagery, the Fractional Snow Cover
classes were grouped into ten classes of 10% differences and one class of 100% snow
cover. The lighter shaded pixels in the MODIS imagery represent higher amounts of
fractional snow cover, while the darker shaded pixels represent less snow cover. Clouds
are represented as cyan, land is represented as brown, and areas with no data are
represented as black.
60
MODIS Fractional Snow Cover (Feb. 23, 2005)
I j 12.5 km AMSR-E
D
Fractional Snow Cover
25 km AMSR-E & SSM/I
. I
•
•
•
Si
H
Pi«s!
\\
B
•
D
•
•
•
•
•
•
•
0%-
9%
10%- 19%
20%- 29%
30% - 39%
40%- 49%
50%- 59%
60% - 69%
70%- 79%
80%- 89%
90%- 99%
100%
Land
Cloud
No Data
Figure 5.22: AMSR-E and SSM/I pixels on MODIS Fractional Snow Cover background
(original in colour)
For this research, it was imperative that the amount of ground snow cover and non-snow
cover was measurable. Unfortunately, it could not be conclusively determined from
several of the MODIS images whether or not there was snow cover under the clouded
areas, and this posed a significant limitation with respect to the analyses of many of the
coincident passive microwave footprints, particularly those that were found to have
considerable cloud cover across the MODIS images. This was not unexpected, as cloudy
and overcast skies were observed during the first two days of the 2005 field campaign,
February 21 st and 22nd, respectively. Nonetheless, spatial analyses were able to be
performed on the MODIS data sets that were not classified with complete cloud cover.
61
For graphing purposes, pixels classified as cloud on the MODIS imagery were assigned a
SWE of 0 mm. Figures 5.23 through 5.50 present the analyses of the MODIS Fractional
Snow Cover imagery.
5.2.6
25 km SSM/I SWE Weighted by MODIS Snow Fractions
Figure 5.23 shows the unweighted (left) and weighted (right) 25 km SSM/I pixel from the
afternoon orbit of February 21, 2005. The MODIS Fractional Snow Cover was largely
classified as cloud cover in this scene. All of the thirteen coincident ground sampling
sites were located under cloud covered pixels. As shown in the image on the left of
Figure 5.23, eight of the thirteen (62%) unweighted sites were within the accuracy of the
MSC algorithm. However, as illustrated in the image on the right, when the weighted
algorithm was applied, a degree of uncertainty was found. The image on the right shows
an additional class for the sampling sites. The additional class, shown as solid yellow
circles, represents sites with an unknown equivalency. Thus, these sites are referred to as
"Unknown Sites." Also shown in Figure 5.23, and all subsequent figures illustrating
MODIS imagery, are the sampling site labels. The sites are labeled by name, with their
corresponding ground SWE measurements in brackets, and where appropriate, the
MODIS Fractional Snow Cover classifications are labeled below the site names and SWE
measurements.
Although all of the sampling sites were under cloud cover, not all of the sites had a
degree of uncertainty. For example, sites D2, D3b, D6 and 603S were still considered to
be within the accuracy of the weighted algorithm because their measured SWE values
would have been within +/- 20 mm SWE of the weighted algorithm regardless of
62
whether the MODIS Fractional Snow Cover was 0% or 100%. Similarly, sites Dl, Fl, Gl
and 603 T were still considered non-equivalent, as their SWE measurements would have
been outside the accuracy of the weighted algorithm regardless of whether the fractional
snow cover was 0% or 100%.
Figure 5.23: Coincident SSM/125 km Pixel 1 MODIS Fractional Snow Cover
for February 21, 2005 P.M. (unweighted on left - weighted on right)
(original in colour)
Due to the extensive cloud cover in the MODIS scene, it was expected that the 0 mm
SWE values assigned to the cloud pixels would significantly bias the weighting of the
estimated SSM/I SWE. Figure 5.24 shows that indeed, the weighted SWE values of 0
mm (represented by the dotted green line) were significantly lower than the ground SWE
measurements (black line) for most of the sampling sites. Figure 5.24 also illustrates that
it was impossible to definitively quantify the difference between the results of the
unweighted (red line) and weighted algorithms (dotted green line) because of the
uncertainty from all of the sampling sites being located under cloud cover. The grey lines
in the graph represent the upper and lower limits of the +/- 20 mm accuracy threshold.
63
SSM/I (25 km) Pixel 1, February 21, 2005 P.M.
SWE Estimate vs. Weighted SWE Estimate
140.0
T
Sampling Site
Figure 5.24: SSM/I 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph
for February 21, 2005 P.M.
(original in colour)
Figure 5.25 presents the results of the 25 km SSM/I pixel from the morning orbit of
February 22, 2005. The MODIS Fractional Snow Cover was entirely cloud covered in
this scene. As such, all seven coincident ground sampling sites were located under cloud
covered pixels. Only two of the seven (29%) unweighted sites were within the accuracy
of the MSC algorithm. Similar to the afternoon orbit from February 21, 2005, and
illustrated in the image on the right in Figure 5.25, when the weighted algorithm was
applied, five of the seven (71%) sites were found to have an unknown equivalency. Thus,
it was again impossible to analyze whether or not the weighted algorithm was an
improvement over the unweighted SSM/I SWE estimate.
64
Figure 5.25: Coincident SSM/125 km Pixel 1 MODIS Fractional Snow Cover
for February 22, 2005 A.M. (unweighted on left - weighted on right)
(original in colour)
Figure 5.26 again illustrates the bias of the cloud covered pixels with 0 mm SWE for the
weighted estimate, as well as the difficulty in quantifying the difference between
weighted and unweighted SWE estimates.
65
SSM/I (25 km) Pixel 1, February 22, 2005 A.M.
SWE Estimate vs. Weighted SWE Estimate
Sampling Site
Figure 5.26: SSM/I 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph
for February 22, 2005 A.M.
(original in colour)
The results of the 25 km SSM/I pixel from the afternoon orbit of February 22, 2005 are
shown in Figure 5.27. The MODIS Fractional Snow Cover was again entirely cloud
covered in this image. Thus, all seven coincident ground sampling sites were again
located under cloud covered pixels. Five of the seven (71%) unweighted sites were within
the accuracy of the MSC algorithm. Similar to the previous examples, the cloud cover
again resulted in a large degree of uncertainty. All seven (100%) sites were found to have
an unknown equivalency when the weighted algorithm was applied, as illustrated in the
image on the right.
66
Figure 5.27: Coincident SSM/125 km Pixel 1 MODIS Fractional Snow Cover
for February 22, 2005 P.M. (unweighted on left - weighted on right)
(original in colour)
As in the previous examples, Figure 5.28 again illustrates the bias of the 0 mm SWE for
the weighted estimates on the cloud covered pixels, and again highlights the difficulty in
quantifying the difference between weighted and unweighted SWE estimates.
67
SSM/I (25 km) Pixel 1, February 22, 2005 P.M.
SWE Estimate vs. Weighted SWE Estimate
140.0
• — S W E Estimate (mm)
130.0
—Weighted SWE Estimate (rr
120.0
^ G r o u n d SWE (mm)
110.0
^—Lower 20 mm Threshold
^ • U p p e r 20 mm Threshold
_ 100.0
i.
90.0
I
80.0
Sampling Site
Figure 5.28: SSM/I 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph
for February 22, 2005 P.M.
(original in colour)
Figure 5.29 presents the results from the 25 km SSM/I pixel from the morning orbit of
February 23, 2005. The MODIS Fractional Snow Cover was again significantly cloud
covered, but six of the eight (75%) sites in this scene were not located under the cloud
covered pixels. As shown in the left image of Figure 5.29, five of the eight (63%)
unweighted sites were found to be within the accuracy of the MSC algorithm.
Interestingly, two of the three sites considered non-equivalent, had a MODIS Fractional
Snow Cover below 80%. Non-equivalent sites E3 and E5 had snow cover classifications
of 78% and 72%, respectively. The third non-equivalent site was located under a cloud
pixel. Of the five sites that were within +/- 20 mm SWE, three were found to have 100%
snow cover, one had 89% snow cover and the other was under a cloud pixel. The
weighted algorithm did not visually appear to improve the SSM/I SWE estimates, as
68
there were no changes to the equivalent and non-equivalent sites, with the exceptions of
sites F3 and G5, which had their classifications change to unknown because they were
coincident with cloud pixels.
Figure 5.29: Coincident SSM/I 25 km Pixel 1 MODIS Fractional Snow Cover
for February 23, 2005 A.M. (unweighted on left - weighted on right)
(original in colour)
Unlike the previous SSM/I examples, Figure 5.30 illustrates that the weighted algorithm
did improve the SSM/I SWE estimates for the sites that were found to have a fractional
snow cover in the MODIS scene. The weighted SWE estimate (dotted green line) is
closer to the ground measured SWE (black line) than the SSM/I SWE estimate (red line)
for sites E2, E3 and E5. Further, although site E5 was not considered equivalent when the
SSM/I estimate was weighted, it did improve the estimate by decreasing it 11.3 mm, a
value within 0.3 mm of the accuracy of the MSC algorithm. Although the weighted
algorithm did improve the accuracy of the SSM/I estimates for fractional snow covered
pixels, caution should be exercised for these analyses, as there was only one coincident
69
sampling site, representative of a 30 m area, for each 500 m MODIS Fractional Snow
Cover pixel.
SSM/I (25 km) Pixel 1, February 23, 2005 A.M.
SWE Estimate vs. Weighted SWE Estimate
Sampling Site
Figure 5.30: SSM/I 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005 A.M.
(original in colour)
Results of the final 25 km SSM/I footprint, from the afternoon orbit of February 23,
2005, are shown in Figure 5.31. The MODIS Fractional Snow Cover was again mostly
cloud covered. The left image shows that five of the seven (71%) sampling sites were
within the accuracy of the SSM/I estimate. Two of the five equivalent sites, F2 and 12,
were coincident with fractional snow covers of 77% and 93%, respectively. However, the
other three equivalent sites were coincident with cloud pixels. The two non-equivalent
sites, G3 and G4, were coincident with fractional snow covers of 70% and 64%,
respectively. Interestingly, the ground SWE measurement for site G3 was 0 mm, while
70
the ground SWE measurement for site G4 was just 1.3 mm. The image on the right shows
that the weighted algorithm did appear to improve the SSM/I estimates, as the two sites
that were non-equivalent in the unweighted image were shown as equivalent in the
weighted image. None of the sites were found to be non-equivalent in the weighted
image, but two of the three sites (G2 and H2) that were coincident with cloud pixels had
their classifications change from equivalent to unknown.
Figure 5.31: Coincident SSM/I 25 km Pixel 1 MODIS Fractional Snow Cover
for February 23, 2005 P.M. (unweighted on left - weighted on right)
(original in colour)
Similar to the results from the morning orbit of February 23, 2005, as illustrated in Figure
5.30, the weighted algorithm did improve the accuracy of the SSM/I estimates for
MODIS Fractional Snow Cover pixels. Figure 5.32 shows an improvement of the
weighted algorithm for sites G3, G4 and 12. The results of the weighted algorithm for
sites G3 and G4 appear to be significant, in that the weighted estimate was within the
accuracy of the MSC algorithm, whereas the unweighted estimate was not. Due to the
fact these two sites had little to no ground-measured SWE, it appears that the weighted
71
algorithm works particularly well over areas of considerable patchy snow cover. As
previously mentioned however, caution should be exercised for these analyses, as there is
only one sampling site that is spatially representative of a 500 m MODIS pixel.
SSM/I (25 km) Pixel 1, February 23, 2005 P.M.
SWE Estimate vs. Weighted SWE Estimate
140.0 -I
^ ^ — S W E Estimate (mm)
130.0
^
120.0
^ ^ ^ G r o u n d SWE (mm)
—Weighted SWE Estimate (mm)
^ " • " U p p e r 20 mm Threshold
110.0 -
^ " • ^ L o w e r 20 mm Threshold
_ 100.0
E
i.
90.0 •s
80.0 •
>
I
70.0 •
ill
te
n
60.0 •
*
50.0 - •
=
40.0
>^
30.0 -
X^
fc
v
^
5
^
^
>^
—
^
^
^1
20.0
"». ^~" «*£ZS.
10.0
tf
&
^s^-
<s>
&
*'
#
^
o
Sampling Site
Figure 5.32: SSM/I 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005 P.M.
(original in colour)
5.2.7
25 km AMSR-E SWE Weighted by MODIS Snow Fractions
Figure 5.33 presents the 25 km AMSR-E pixel from the orbit of February 21, 2005. As
with the SSM/I estimate for the afternoon orbit for this date, the AMSR-E footprint was
significantly cloud covered and all of the thirteen (100%) sampling sites were coincident
with cloud pixels. The left image in Figure 5.33 shows that three of the thirteen (23%)
sites were equivalent to the 25 km AMSR-E SWE estimate, while the remaining ten
(77%) sites were non-equivalent. Similar to the analysis of the SSM/I estimate from
72
February 21, 2005, the right image in Figure 5.33 shows that the weighted algorithm
created a large amount of uncertainty, as twelve of the thirteen (92%) sites were
classified as unknown.
Figure 5.33: Coincident AMSR-E 25 km Pixel 1 MODIS Fractional Snow Cover
for February 21, 2005 (unweighted on left - weighted on right)
(original in colour)
With the exception of the AMSR-E SWE estimate (red line), Figure 5.34 is identical to
Figure 5.24. The graph shows that the biased weighted SWE values of 0 mm (dotted
green line) were significantly lower than the ground SWE measurements (black line) for
most of the sampling sites. Figure 5.34 also illustrates that it was impossible to quantify
the difference between the results of the unweighted and weighted algorithms, because of
the uncertainty from all of the sampling sites being coincident with cloud pixels.
73
Figure 5.34: AMSR-E 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph
for February 21, 2005
(original in colour)
Results of the AMSR-E footprint from the orbit of February 22, 2005 are presented in
Figure 5.35. As with the footprints of the SSM/I estimates collected from the morning
and afternoon orbits of February 22, 2005, the footprint of the AMSR-E estimate from
this date was also entirely classified as cloud by the MODIS Fractional Snow Cover. The
image on the left shows that only two of the fourteen (14%) coincident sites (603 G and
6031) were found to be equivalent with the AMSR-E estimate, while the other twelve
(86%) were non-equivalent. The image on the right shows that when the weighted
algorithm was applied, all fourteen (100%) sites were classified as unknown.
74
Figure 5.35: Coincident AMSR-E 25 km Pixel 1 MODIS Fractional Snow Cover
for February 22, 2005 (unweighted on left - weighted on right)
(original in colour)
Similar to the SSM/I estimates collected from the morning and afternoon orbits of
February 22, 2005, Figure 5.36 again illustrates the bias of the 0 mm SWE for the
weighted estimates on the cloud covered pixels and the difficulty in analyzing the results
of the weighted algorithm.
75
Figure 5.36: AMSR-E 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph
for February 22, 2005
(original in colour)
Results of the third, and final, 25 km AMSR-E footprint, from the orbit of February 23,
2005, are shown in Figure 5.37. The left image shows that five of the fifteen (33%)
sampling sites were within the accuracy of the AMSR-E estimate. Only one of the five
equivalent sites, E2, was coincident with a fractional snow cover. Site E2 had fractional
snow cover of 89%. Three of the other four equivalent sites were coincident with 100%>
snow cover, and the fifth equivalent site was coincident with a cloud pixel. Of the ten
non-equivalent sites, four were coincident with cloud pixels and one was coincident with
a 93% snow cover. Interestingly, the remaining five were coincident with fractional snow
covers ranging from 64% to 78%. The image on the right shows that the weighted
algorithm did not visually appear to significantly improve the AMSR-E estimates, as only
one of the ten sites (F2) shown as non-equivalent in the unweighted image was shown as
76
equivalent in the weighted image. The five sites coincident with cloud pixels were all
changed to unknown, and the five remaining non-equivalent sites coincident with
fractional snow cover were unchanged.
Figure 5.37: Coincident AMSR-E 25 km Pixel 1 MODIS Fractional Snow Cover
for February 23, 2005 (unweighted on left - weighted on right)
(original in colour)
Similar to the results from the SSM/I morning and afternoon orbits of February 23, 2005,
as illustrated in Figures 5.30 and 5.32, the weighted algorithm did improve the accuracy
of the AMSR-E estimates for MODIS Fractional Snow Cover pixels. Figure 5.38 shows
an improvement of the weighted algorithm for sites E2, E3, E5, F2, G3, G4 and 12. The
results of the weighted algorithm for site F2 appears to be significant, as the weighted
estimate was within the accuracy of the MSC algorithm, while the unweighted estimate
was not. Although none of the other fractional snow covered sites were found to be
improved to a value within the MSC algorithm accuracy when the weighted algorithm
was applied, they were indeed improved.
77
Figure 5.38: AMSR-E 25 km Pixel 1 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005
(original in colour)
5.2.8
12.5 km AMSR-E SWE Weighted by MODIS Snow Fractions
As illustrated with the MODIS scenes that were coincident with the 25 km SSM/I and 25
km AMSR-E footprints, the pixels classified as cloud were impossible to quantifiably
analyze. Due to the difficulty in analyzing cloud covered scenes, all of the 12.5 km
AMSR-E footprints coincident with MODIS scenes containing 100% cloud classified
sampling sites have been omitted from the 12.5 km AMSR-E analysis. Thus, only the
February 23, 2005 imagery is presented in this section.
The first 12.5 km AMSR-E footprint (Pixel 1) from the orbit of February 23, 2005 is
presented in Figure 5.39. As in the 25 km AMSR-E footprint, the pixels in the MODIS
scene were largely classified as cloud. Only one coincident sampling site was collected
78
for this date, site F2. This site was coincident with a MODIS pixel having a 77% snow
cover. The image on the left shows that the site was within the accuracy of the 12.5 km
AMSR-E SWE estimate, while a visual assessment of the image on the right suggests that
the weighted algorithm did not improve the estimate.
Figure 5.39: Coincident AMSR-E 12.5 km Pixel 1 MODIS Fractional Snow Cover
for February 23, 2005 (unweighted on left - weighted on right)
(original in colour)
Figure 5.40 presents the results of the 12.5 km AMSR-E Pixel 1 data. For graphing
purposes, site F2 is duplicated in the graph. The graph illustrates that although the
AMSR-E SWE estimate was within the accuracy of the MSC algorithm, the weighted
algorithm did indeed improve the AMSR-E estimate, by decreasing the estimate by
approximately 7.0 mm, a value nearly identical to the ground measurement.
79
AMSR-E (12.5 km) Pixel 1, February 23, 2005
SWE Estimate vs. Weighted SWE Estimate
140.0
- S W E Estimate (mm)
130.0
-Weighted SWE Estimate (mrr
120.0
• Ground SWE (mm)
-Upper 20 mm Threshold
110.0
-Lower 20 mm Threshold
•£• 100.0
E
r 9o.o
70.0
60.0
50.0
40.0
30.0
20.0
10.0
0.0
^
^
Sampling Site
Figure 5.40: AMSR-E 12.5 km Pixel 1 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005
(original in colour)
Figure 5.41 shows results of the second unweighted and weighted 12.5 km AMSR-E
pixels (Pixel 2). A total of six sampling sites were coincident with this AMSR-E footprint
and MODIS scene. The image on the left shows that three of the six sites were classified
with fractional snow, two sites were classified with 100% snow cover and one site was
classified as cloud. Further, three of the sites were found to be equivalent, and each of the
equivalent sites was classified with a snow cover greater than 89%. Of the three nonequivalent sites, two were classified with snow covers of 70% and 78%, sites G3 and E3,
respectively, while the third non-equivalent site was classified as cloud.
The image on the right suggests an improvement of the weighted algorithm by showing a
higher number of equivalent sites than the unweighted image on the left. The three sites
80
found to be equivalent with the unweighted AMSR-E estimate shown on the left,
remained equivalent when the weighted algorithm was applied, while a fourth site (E3)
was also found to be equivalent with the weighted algorithm.
Figure 5.41: Coincident AMSR-E 12.5 km Pixel 2 MODIS Fractional Snow Cover
for February 23, 2005 (unweighted on left - weighted on right)
(original in colour)
Figure 5.42 illustrates the improvement of the weighted algorithm for sites with fractional
snow cover of less than 80% (i.e., sites E3 and G3). Although site G3, which is plotted on
the furthest right side of the graph, was not found to be equivalent when the weighted
algorithm was applied, the weighted estimate was decreased to a value much closer to the
ground measurement.
81
Figure 5.42: AMSR-E 12.5 km Pixel 2 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005
(original in colour)
Results of the third 12.5 km AMSR-E footprint (Pixel 3) are shown in Figure 5.43. Only
two sampling sites were coincident with this scene. One of the sites (F5) was classified as
100% snow cover, while the other (E5) was classified as 72% snow cover. The results for
Pixel 3 were similar to the Pixel 2 results, where the 100% snow cover was found to be
equivalent to the AMSR-E SWE estimate and the 72% snow cover was non-equivalent. A
visual comparison of the unweighted (left) and weighted (right) images indicates that the
weighted algorithm did not improve the accuracy of the AMSR-E estimates for either of
the sites.
82
Figure 5.43: Coincident AMSR-E 12.5 km Pixel 3 MODIS Fractional Snow Cover
for February 23, 2005 (unweighted on left - weighted on right)
(original in colour)
As shown in Figure 5.44, the weighted algorithm did improve the accuracy of the AMSRE estimate for sampling site E5, the site with a fractional snow cover of 72%. Although
this site, which is plotted on the left side of the graph, was not found to be equivalent
when the weighted algorithm was applied, the weighted estimate was again decreased to
a value much closer to the ground measurement than the unweighted estimate.
83
AMSR-E (12.5 km) Pixel 3, February 23, 2005
SWE Estimate vs. Weighted SWE Estimate
140.0
^ S W E Estimate (mm)
130.0
— Weighted SWE Estimate (mm
120.0
^ G r o u n d SWE (mm)
^ " U p p e r 20 mm Threshold
110.0
•••Lower 20 mm Threshold
•g- 100.0
E
r 9o.o
c
g
80.0
S
70.0
Sampling Site
Figure 5.44: AMSR-E 12.5 km Pixel 3 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005
(original in colour)
Figure 5.45 presents the results of AMSR-E 12.5 km Pixel 4. The left image shows that
the unweighted estimates were within the accuracy of the MSC algorithm for all three
coincident sampling sites. Two of the sites were coincident with cloud pixels, while the
third site (12) was coincident with a pixel classified as 99% snow cover. The image on the
right indicates that the weighted algorithm produced a similar result to the unweighted
SWE estimate for a MODIS pixel classified with 99% fractional snow cover.
84
Figure 5.45: Coincident AMSR-E 12.5 km Pixel 4 MODIS Fractional Snow Cover
for February 23, 2005 (unweighted on left - weighted on right)
(original in colour)
Similar to the results from Pixel 2 and Pixel 3, Figure 5.46 shows that the weighted
estimate was within the AMSR-E estimate for sampling sites with fractional snow cover
near 100%. Although the graph in Figure 5.46 does not illustrate a significant
improvement in the weighted algorithm for site 12, a significant improvement was not
expected, as the weighted SWE estimate was only decreased by 0.4 mm because of the
99% snow cover.
85
Figure 5.46: AMSR-E 12.5 km Pixel 4 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005
(original in colour)
Figure 5.47 presents the results of the fifth AMSR-E 12.5 km pixel (Pixel 5). While four
sampling sites were coincident with this footprint, three of the four were classified as
cloud. The only site classified with fractional snow was site G4, which was classified
with 64% snow cover and was found to be non-equivalent. The right image in Figure
5.47 does not illustrate a visual improvement in the weighted estimate of site G4, as the
site was still classified as non-equivalent.
86
Figure 5.47: Coincident AMSR-E 12.5 km Pixel 5 MODIS Fractional Snow Cover
for February 23, 2005 (unweighted on left - weighted on right)
(original in colour)
Although improvement in the weighted estimate for site G4 was not visually evident in
Figure 5.47, the graph in Figure 5.48 does show improvement in the weighted algorithm.
As illustrated in the graph, the weighted estimate for site G4 was much closer to the
ground measurement than the unweighted AMSR-E estimate.
87
Figure 5.48: AMSR-E 12.5 km Pixel 5 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005
(original in colour)
Results of the sixth 12.5 km AMSR-E footprint (Pixel 6) are presented in Figure 5.49. A
total of five coincident sampling sites were located in this AMSR-E footprint. The left
image shows that four of the five sites were within the accuracy of the MSC algorithm,
while two of these four sites were coincident with cloud pixels. Of the two equivalent
sites classified with fractional snow, one (F6) was classified with 100% snow cover and
the other (H6) was classified with 55% snow cover. The lone non-equivalent site (H7)
was classified with an 86% snow cover. The image on the right shows no visually
apparent improvement in the weighted SWE estimates, as the sites classified with
fractional snow covers (F6, H6 and H7) are classified with the same equivalencies as in
the unweighted image shown on the left.
88
Figure 5.49: Coincident AMSR-E 12.5 km Pixel 6 MODIS Fractional Snow Cover
for February 23, 2005 (unweighted on left - weighted on right)
(original in colour)
As in the previous 12.5 km AMSR-E examples, the weighted estimates for the sites
classified with fractional snow covers were found to be improved over the unweighted
estimates. The graph in Figure 5.50 illustrates that although the equivalencies of sites H6
(55% snow cover) and H7 (86% snow cover) were unaffected by the weighted algorithm,
the weighted estimates for these sites were indeed improved over the AMSR-E estimates.
89
AMSR-E (12.5 km) Pixel 6, February 23, 2005
SWE Estimate vs. Weighted SWE Estimate
140.0
- S W E Estimate (mm)
130.0
• Weighted SWE Estimate (mm)
120.0
• Ground SWE (mm)
• Upper20mm Threshold
110.0
• Lower 20 mm Threshold
-£ 100.0
£
r 90.0
c
g
80.0
1
70.0
Sampling Site
Figure 5.50: AMSR-E 12.5 km Pixel 6 SWE vs. MODIS-Weighted SWE graph
for February 23, 2005
(original in colour)
5.3
Spatial Analysis Validation Results
In order to validate the findings of the fractional snow cover analyses on the 2005 data
sets, a second field campaign was undertaken on March 2, 2008. The spatial extent of the
2008 field campaign was less extensive than in 2005 and focused on the geographic area
within AMSR-E 12.5 km Pixels 1 and 2 (Figures 5.5 and 5.22). All data collection and
data processing methods and techniques used for the 2005 study were reproduced for the
2008 field data and remote sensing data sets. Field data were collected from a total of 18
sampling sites, nine sites that were well-distributed and spatially representative of each of
the two AMSR-E 12.5 km pixels. Clear skies were observed from all but one of the
sampling sites, so cloud cover was not expected to impact the image analyses.
90
53ot
Fractional Snow Cover Validation from TM Imagery
Figure 5.51 shows the results of the unweighted (left) and weighted (right) images for the
first 12.5 km AMSR-E validation pixel (Pixel 1), as collected from the orbit of March 2,
2008. The TM-derived fractional snow cover of this footprint was 97%. Ground-truth
observations collected from nine sampling sites within the pixel yielded an average insitu SWE measurement of 29.0 mm. Of the nine sites, six (67%) were found to be within
the MSC SWE algorithm accuracy of +/- 20 mm. The unweighted AMSR-E SWE
estimate was 40.0 mm, which was within the estimated accuracy of the MSC SWE
algorithm. The weighted SWE estimate was 38.8 mm, a value that was closer to the
ground measurement.
Figure 5.51: AMSR-E 12.5 km Pixel 1 TM Fractional Snow Cover for March 2, 2008
(unweighted on left - weighted on right)
(original in colour)
91
Although an improvement in the weighted estimate was not visually apparent in Figure
5.51, the graph in Figure 5.52 illustrates that the weighted estimate was indeed closer to
the ground measurement than the AMSR-E estimate.
A M S R - E (12.5 k m ) Pixel 1 , March 2, 2 0 0 8
S W E Estimate vs. Weighted S W E Estimate
70.0 -i
™ » — SWE Estimate (mm)
••» •Weighted SWE Estimate (mm)
^ • • " M e a n Ground SWE (mm)
™ ^ i » U p p e r 20 mm Threshold
^ ^ ^ » L o w e r 20 mm Threshold
60.0
50.0
E
+•20 mm Threshold
g
2
>
40.0 •
'5
c
LU
S 30.0 3
1
w 20.0 •
- 20 mm Threshold
10.0
A
0.0 •
Date (Time)
Figure 5.52: AMSR-E 12.5 km Pixel 1 SWE vs. TM-Weighted SWE graph for March 2,2008
(original in colour)
Results from the second AMSR-E 12.km validation pixel (Pixel 2) from March 2, 2008
are shown in Figure 5.53. The TM-derived fractional snow cover of this footprint was
92%. Ground-truth observations collected from the nine sampling sites within this pixel
yielded an average in-situ SWE measurement of 13.8 mm. The unweighted AMSR-E
SWE estimate was 58.2 mm. Interestingly, the satellite value was beyond the +/- 20 mm
accuracy threshold of all of the ground SWE values. The weighted SWE estimate was
53.5 mm, a value that was closer to, but still significantly higher than, the ground
measurement.
92
Figure 5.53: AMSR-E 12.5 km Pixel 2 TM Fractional Snow Cover for March 2, 2008
(unweighted on left - weighted on right)
(original in colour)
As in the validation results from Pixel 1, an improvement in the weighted estimate was
not visually apparent in Figure 5.53. However, the graph in Figure 5.54 again illustrates
that the weighted estimate was closer to the ground measurement than the AMSR-E
estimate.
93
AMSR-E (12.5 km) Pixel 2, March 2, 2008
S W E Estimate vs. Weighted S W E Estimate
70.0
^ ^ — SWE Estimate (mm)
—
—Weighted SWE Estimate (mm)
i
^ ^ ^ • M e a n Ground SWE (mm)
60.0
•••••••Upper 20 mm Threshold
••••••» Lower 20 mm Threshold
50.0
£
40.0 •
>
"5
c
+ 20 mm Threshold
01
S 30.0 •
(0
s
S
o
•
<0 20.0
- 20 mm Threshold
10.0
0.0 •
Date (Time)
Figure 5.54: AMSR-E 12.5 km Pixel 2 SWE vs. TM-Weighted SWE graph
for March 2, 2008
(original in colour)
5.3.2
Fractional Snow Cover Validation from MODIS Imagery
Figure 5.55 presents the MODIS validation results of AMSR-E 12.5 km Pixel 1 from the
orbit of March 2, 2008. Of the nine sampling sites, five were coincident with cloud pixels
and four were coincident with 100% snow cover. While many 100% snow covered pixels
were expected, due to the 97% snow cover found in the coincident TM image for this
date, numerous cloud-classified pixels were not expected, as clear skies were observed at
nearly all of the sampling sites during the field data collection. Due to all of the sites
being classified as cloud or 100% snow cover, it was impossible to analyze the impact of
fractional snow cover for this scene.
94
AWISR-E 12.5km Pixel 1
AMSR-E 12.5km Pixel 1
^^^^^^^^^
mm^^ ^
W^
H 125.21 ^ L H 1 MODIS (March 2,2008)
io'o%*
^ H 1 Fractional Snow Cover
G (37.6)
• 100%
1•
1•
1•
1•
1•
1
1 c•
1D
1(4.6)1 1 C
Cloud 1 1 •
1
1a
•
1•
10%
19%
20%
29%
30%
A (46.1)
# Cloud
1 MODIS (March 2, 2008)
1 Fractional Snow Cover
H (25.2)
00% »
0%- 9%
1 H 10%1 • 20%- 1B%
29%
1 •
1 • 30%- 39%
40%19%
1 •
1 • 50%- 59%
1 E 60%- 69%
79%
1 LVi 70%B0%- 69%
0%- 9%
39%
00%
19%
50%
59%
60%
69%
70%
79%
80%
90%
G (37.6)
• 100%
A (46.1)
' ) Cloud
B9%
(4.6)1
99%
-Cloud •
100%
Cloud
(38.1)
Cloud
CI16.1)
• 100%
D (21.6)
• 100%
100%
1U
1•
Cloud
F(51.6)
Cloud
1•
No Data
E(17.7)
• Cloud
90%- 99%
Land
Land
F (S1.6)
Cloud £
1: ' '
1 n1
1
1•
B (38.1|
Cloud
1
W^
No Data
E(17.7)
D (21.6)
• 100%
C (16.1)
• 100%
AMSR-E SWE = 40.0 r
Figure 5.55: Coincident AMSR-E 12.5 km Pixel 1 MODIS Fractional Snow Cover
for March 2, 2008 (unweighted on left - weighted on right)
(original in colour)
The graph in Figure 5.56 illustrates the difficulty in analyzing the data from this scene.
A M S R - E (12.5 km) Pixel 1 , M a r c h 2, 2 0 0 8
S W E Estimate v s . W e i g h t e d S W E Estimate
Sampling Site
Figure 5.56: AMSR-E 12.5 km Pixel 1 SWE vs. MODIS-Weighted SWE graph
for March 2, 2008
(original in colour)
95
Results from the second AMSR-E 12.km MODIS validation pixel (Pixel 2) from March
2, 2008 are shown in Figure 5.57. As in the MODIS validation Pixel 1, all nine of the
sampling sites were coincident with either 100% snow cover or cloud pixels, which again
made it impossible to analyze the impact of fractional snow cover.
Figure 5.57: Coincident AMSR-E 12.5 km Pixel 2 MODIS Fractional Snow Cover
for March 2, 2008 (unweighted on left - weighted on right)
(original in colour)
Similar to Figure 5.56, the graph in Figure 5.58 illustrates the difficulty in analyzing the
data from this scene.
96
AMSR-E (12.5 km) Pixel 2, March 2, 2008
SWE Estimate vs. Weighted SWE Estimate
140.0 -I
— " — S W E Estimate (mm)
130.0 •
—
—Weighted SWE Estimate (mm)
«^^»GroundSWE(mm)
^ " • " * U p p e r 2 0 mm Threshold
110.0
^ ^ ^ L o w e r 20 mm Threshold
-£• 100.0
E
r
9o.o
g
80.0
S
70.0
c
V
|
60.0
S
0
c
50.0 -
m
40.0
\
k
'
\
***»^
\
'
^
\
~ - " ^ i
"*w
x^,^
\
/
^ \ ^
^^""—•"v^
>^
*
>*. y^
x*.
^ ^ / \
_-^-—V ^ ^ V'"*^"--^
\
20.0
10.0 -
"—-^^
\
^^^^
•Jr
^
v
x.>
J^'
\ '
y^
v>l--- J »f~**^^
^-
•£•
O
<2
^ ^
x.
O-
«-
Sampling Site
Figure 5.58: AMSR-E 12.5 km Pixel 2 SWE vs. MODIS-Weighted SWE graph
for March 2, 2008
(original in colour)
5.4
Results Discussion
5.4.1
Fractional Snow Cover Derived from TM Imagery
Results from the Snow/No-Snow Landsat TM classifications showed that the accuracy of
the MSC algorithm was improved by weighting the remotely sensed SWE estimates by
the fractional snow cover across the passive microwave footprints. In total, there were
twenty-one passive microwave derived SWE estimates analyzed from the 2005
campaign. Tables 13 through 15 show the accuracies of the unweighted and weighted
SWE estimates for each of the five coincident SSM/I 25 km footprints, three AMSR-E 25
km footprints, and thirteen AMSR-E 12.5 km footprints, respectively. In these tables,
"SWE" refers to unweighted estimates, "SWEW" refers to weighted estimates,
97
"Equivalent" refers to estimates that were found to be within the +/- 20 mm accuracy of
MSC algorithm, and "Overestimated" and "Underestimated" refer to estimates found to
be outside of the MSC algorithm accuracy.
Table 13 shows that all five of the unweighted SSM/I 25 km SWE estimates were within
the accuracy of the MSC algorithm. Also shown in this table is that all of the weighted
SWE estimates were also found to be within +/- 20 mm SWE, and two of the five (40%)
weighted estimates were improvements over the unweighted estimates.
Table 14 illustrates that all three of the unweighted AMSR-E 25 km estimates were
overestimated, and thus, outside the accuracy of the MSC algorithm. Table 14 also shows
that all three (100%) of the weighted SWE estimates were improved to within +/- 20 mm
of SWE.
98
Coincident
Pixel
59%
59%
59%
59%
59%
Percent
Snow Cover
5/5
Equivalent
SWE
0/5
Overestimated
SWE
0/5
Underestimated
SWE
5/5
Equivalent
SWEW
Feb 21/05
Feb 22/05
Feb 23/05
TOTAL
Date/Orbit
1
1
1
Coincident
Pixel
59%
59%
59%
Percent
Snow Cover
0/3
Equivalent
SWE
3/3
•
•
•
Overestimated
SWE
0/3
Underestimated
SWE
3/3
•
•
•
Equivalent
SWEW
Table 14: 25 km AMSR-E SWE estimates integrated with TM imagery (• denotes "YES")
TOTAL
Feb21P.M./05
Feb 22 A.M./05
Feb 22 P.M./05
Feb 23 A.M./05
Feb 23 P.M./05
Date/Orbit
Table 13: 25 km SSM/I SWE estimates integrated with TM imagery (• denotes "YES")
3/3
•
•
•
Improved
SWEW
2/5
•
•
Improved
SWEW
The accuracies for the weighted and unweighted AMSR-E 12.5 km SWE estimates are
presented in Table 15. This table shows that while nine of the thirteen (69%) unweighted
estimates from the 2005 campaign were within the accuracy of the MSC algorithm, four
of the thirteen (31%) were overestimated. Table 15 also shows that although the weighted
algorithm improved only nine of the thirteen (69%) unweighted 2005 estimates, that
eleven of the thirteen (85%) weighted estimates were within +/- 20 mm.
Analyses of the 2008 campaign are also shown in Table 15, and results similar to the
2005 campaign were found when the weighted SWE algorithm was validated on the 2008
data sets. A total of two passive microwave derived SWE estimates were analyzed from
the 2008 validation campaign, both coincident AMSR-E 12.5 km footprints. Of the two
footprints, one (50%) was found to be overestimated (Pixel 2), while the other (Pixel 1)
was found to be within the MSC algorithm threshold. Although the weighted SWE value
did not improve the overestimated AMSR-E SWE estimate to be within +/- 20 mm, it did
lower the estimated value to be in closer agreement with the ground measurement.
Another reference to Table 15 shows that when the accuracies of the AMSR-E 12.5 km
SWE estimates from the 2008 campaign were included with the 2005 analyses, 67% (10
of 15) of the unweighted estimates were found to be within +/- 20 mm, while 33% (5 of
15) of the unweighted estimates were found to be overestimated. Analyses of the SWE
estimates from both campaigns showed that the weighted algorithm improved 73 % (11 of
15) of the unweighted estimates, and 80% (12 of 15) of the weighted estimates were
within the accuracy of the MSC algorithm.
100
Table 15: 12. 5 km AMSR-E SWE estimates integrated with TM imagery (• denotes "YES")
Date/Orbit
Coincident
Pixel
Percent
Snow Cover
Equivalent
SWE
Feb 21/05
Feb 22/05
Feb 23/05
Feb 21/05
Feb 23/05
Feb 21/05
Feb 22/05
Feb 23/05
Feb 21/05
Feb 23/05
Feb 23/05
Feb 22/05
Feb 23/05
1
1
1
2
2
3
3
3
4
4
5
6
6
45%
45%
45%
58%
58%
89%
89%
89%
52%
52%
39%
67%
67%
•
•
•
Mar 2/08
Mar 2/08
2005 Total
2008 Total
TOTAL
1
2
97%
92%
•
Overestimated
SWE
Underestimated
SWE
Equivalent
SWEW
Improved
SWEW
0/13
0/2
0/15
11/13
1/2
12/15
9/13
2/2
11/15
•
•
•
•
•
•
•
•
•
•
•
9/13
1/2
10/15
4/13
1/2
5/15
In total, 35% (8 of 23) of all unweighted SWE estimates, from both campaigns, were
inaccurately overestimated (i.e., above the +/- 20 mm MSC algorithm accuracy), while
none (0%) of the unweighted SWE estimates were inaccurately underestimated. All eight
of the overestimated SWE values were associated with the AMSR-E sensor; all three
(100%) AMSR-E 25 km footprints, and five of the fifteen (33%) AMSR-E 12.5 km
footprints. In all eight overestimated SWE values, the weighted algorithm improved the
SWE estimates, and seven of the eight weighted estimates were found to be within the
accuracy of the MSC algorithm. Further, 70% (16 of 23) of the weighted SWE estimates
were found to be in closer agreement with the ground SWE measurements than the
unweighted SWE estimates. The only data sets where the accuracy was not improved
were the three afternoon orbits of the SSM/I sensor, all three orbits of AMSR-E 12.5 km
Pixel 1, and the February 23rd orbit of AMSR-E 12.5 km Pixel 4.
The unweighted SWE estimates from the February 22nd and 23rd, 2005 afternoon SSM/I
orbits can be explained by the weather observations and ground measurements collected
from those afternoon dates, as well as the impact of wet snow conditions on passive
microwave SWE retrievals. Field data collected on the afternoons of the second two dates
were found to have higher air and ground temperatures, and different snow states (moist
as opposed to dry snow), than data collected from the afternoon of the first field
campaign date. A wet snowpack will negatively impact remotely sensed SWE estimates
by saturating the TB signal which causes the SWE algorithm to produce an estimate of 0
mm SWE. Thus, it was expected that the SWE retrievals from the afternoon SSM/I orbits
would be lower than the SWE estimates from the morning orbits, and the lower SWE
102
estimates weighted by the fractional snow cover resulted in further decreased estimates.
Although the weighted SWE algorithm did not improve the accuracy of the SSM/I SWE
estimates from the afternoon orbits, the weighted SWE values were still within the +/- 20
mm threshold.
With respect to the results from AMSR-E 12.5 km Pixel 1, it is assumed that because the
sampling sites were not spatially representative of the AMSR-E footprint, the ground
SWE measurements were not accurately representative of the footprint. Therefore,
performance of the weighted SWE algorithm for this pixel is inconclusive.
As the results from the 2008 study validated the results of the 2005 study, it can be
concluded that the weighted SWE algorithm generated from the iNDSI algorithm using
Landsat TM imagery does improve the unweighted passive microwave SWE estimates
over patchy snow covered areas. Further, it was found that a 92% snow cover can
significantly impact the results of passive microwave derived SWE estimates at 12.5 km
spatial resolution (Figure 5.53). Although caution should be exercised for passive
microwave footprints containing few ground sampling measurements (Figures 5.14 and
5.18), and/or sampling sites that are not well-distributed and spatially unrepresentative of
the footprints (Figure 5.10).
5.4.2
Fractional Snow Cover Derived from MODIS Imagery
Similar to the results from the Landsat TM research, results from the MODIS study
showed that the accuracy of the MSC algorithm was improved by weighting the remotely
sensed SWE estimates by the MODIS Fractional Snow Cover. In total, there were
103
twenty-three passive microwave derived SWE estimates analyzed from the 2005
campaign. Only those sites coincident with MODIS Fractional Snow Cover pixels are
considered in this discussion, as the coincident cloud pixels were impossible to quantify,
and the 100% snow cover pixels produced identical results for weighted and unweighted
SWE estimates. Tables 16 through 18 show the accuracies of the unweighted and
weighted SWE estimates for the coincident SSM/I 25 km, AMSR-E 25 km and AMSR-E
12.5 km footprints, respectively. All of the unweighted SWE estimates were obtained
from the orbits of February 23, 2005, as this was the only date that was found to have
sampling sites that were not coincident with cloud covered pixels. As in Tables 13
through 15, in these tables, "SWE" refers to unweighted estimates, "SWEW" refers to
weighted estimates, "Equivalent" refers to estimates that were found to be within the +/20 mm accuracy of MSC algorithm, and "Overestimated" and "Underestimated" refer to
estimates found to be outside of the MSC algorithm accuracy.
Table 16 shows that only three of the seven (43%) unweighted SSM/I 25 km SWE
estimates were within the accuracy of the MSC algorithm. Table 16 also shows that five
of the seven (71%) weighted SWE estimates were found to be within +/- 20 mm SWE,
and six of the seven (86%) weighted estimates were improvements over the unweighted
estimates.
Table 17 illustrates that only one of the seven (14%) unweighted AMSR-E 25 km
estimates was within the accuracy of the MSC algorithm, and the other six (86%) were
overestimated. Table 17 also illustrates that all seven (100%) of the weighted SWE
estimates were improved, but only two of the seven (29%) were considered equivalent.
104
6/7
Improved
SWEW
•
•
•
Equivalent
SWEW
SV
•4-1
4/7
WD
«
t/5
•
•
•
v
03
U
•v©
%0 sO
0 s ©** 0 s
ON 0 0
CN
oo r-
NP
©^
C-»
r- r-
x®s
0
O
x®
©**
'Tf
N Ps
r- vo
ON
OH
0
m
w
«
C/2
=
«
I
Feb 23 A.M./05
Feb 23 A.M./05
Feb 23 A.M./05
Feb 23 P.M./05
Feb 23 P.M./05
Feb 23 P.M./05
Feb 23 P.M./05
tZ!
Date/Orbit
W
Coincident
Pixel
1
S
0/7
1/7
O
Equivalent
SWE
Q
•
S©
©^
"S©
0 s
>/->
o
•*
CN
W w pq m o
O
"
CN
t-
H
105
\ 0
0 s
O
r-
N©
0s*
•*
vc
N©
©**
f)
o\
>n >/-> i/-> v->
© © o
o
o m m ci
CN
CN
CN
.n xi
(L>
PH
Sampling
Site
fN m
TOTAL
Sampling
Site
.B
vX
0 s
>r> <n i/i
o © ©
ro r«i f )
<D
PH
IT)
VC
s*0
0s-
0 \ oo fN t oo r- r- r-
-B
CN m
CN
CN
CN
-Q .£> x> jz
<L>
m
<L>
P-<
""1 CN
CN
x>
<L> d > <L>
PH
PH
b
f>
"^
(N
W W W H. O
O
"
TOTAL
ss
S.
•
Percent
Snow Cover
0
•
Coincident
Pixel
Q
•
Date/Orbit
C/5
•
3/7
a
Equivalent
SWE
es
Percent
Snow Cover
I.
V
01
•
6/7
ES
1)
T3
•
Overestimated
SWE
0/7
o
7/7
•
Underestimated
SWE
•
5/7
Improved
SWEW
Equivalent
SWEW
•
•
Underestimated
SWE
on
<U
•^
o
c<&
•
5/3
Overestimated
SWE
W
•
•
A/3
00
•
The accuracies for the weighted and unweighted AMSR-E 12.5 km SWE estimates are
presented in Table 18. This table shows that while four of the nine (44%) unweighted
estimates were within the accuracy of the MSC algorithm, the other five (56%) were
overestimated. Table 18 also illustrates that although the weighted algorithm improved
eight of the nine (89%) AMSR-E estimates, that only five (56%) weighted estimates were
within +/- 20 mm.
106
LOl
Sampling
Site
TOTAL
ffiffiOr-jWOtfltTi'-rt
*Tl
*TJ
*Tj
'Tj
'Tj
'Tj
*Tl *"Tl
o'a-crcra'a'a-cro'
O
O
O
O
O
O
O
O
O
o^
x?
p**
x?
5s-
NP
NP
NP
o^
o^
©** p**
•
•
•
•
•
•
Improved
SWEW
•
•
Equivalent
SWEW
•
NP
Overestimated
SWE
•
Underestimated
SWE
6/0
5/9
•
sP
©""•
5/9
6/8
Equivalent
SWE
4/9
•
sP
p^
Percent
Snow Cover
NO
Coincident
Pixel
O N O \ < - f « 4 ^ 0 J K J K J K ) H -
Date/Orbit
*T^
•
•
•
• • • • • •
•
In total, twenty-one of the twenty-three (91%) MODIS-weighted SWE estimates were in
closer agreement with the ground measurements than with the unweighted SWE
estimates. However, only twelve of the twenty-three (52%) MODIS-weighted SWE
values were within the +/- 20 mm accuracy of the MSC algorithm. The relatively low
percentage of sites found to be within the accuracy of the MSC algorithm can be
explained by the fact that only one sampling site, representative of a 30 m area, was
representative of one 500 m MODIS pixel. As a great deal of land cover variability is
possible across the area covered by one MODIS pixel, the characteristics of one sampling
site representative of that pixel will bias the results. For example, although a sampling
site may have been located on a fallow field that contained little to no snow cover, that
sampling site is representative of a 500 m pixel, which may contain numerous stubble
fields, shelter belts or pastures with varying snow covers, and therefore varying snow
characteristics. Therefore, caution should be exercised for passive microwave footprints
containing few ground sampling measurements, and/or sampling sites that are not welldistributed and spatially unrepresentative of the footprints, since this can produce
misleading results.
Unlike the analysis of the TM imagery, cloud cover limited the analysis of a significant
portion of the MODIS data. There were many more pixels classified as cloud in the
MODIS imagery than were visually apparent in the contemporary TM data. Although
clouds are clearly visible in the higher resolution TM imagery, they did not obscure the
surface as much as the MODIS data would suggest. Three possible reasons for this
inconsistency are: i) scaling issues between the 30 m TM imagery and the 500 m MODIS
108
data; ii) incorrectly classified cloud pixels in the MODIS imagery; and iii) the MODIS
cloud mask being triggered by optically-thin cirrus clouds.
First, it takes approximately 275 Landsat TM pixels to cover the same ground area as a
MODIS 500 m footprint. Even if many of the TM pixels are not cloud covered, if the
cloudy area within a MODIS footprint reaches a critical level, the cloud detection
algorithm will mark the MODIS pixel as being cloud-covered. It is impossible to
determine what the critical cloud threshold is to trigger the MODIS cloud mask because it
depends on the cloud thickness (i.e., optically clear vs. opaque), distribution (i.e.,
clumped vs. widely distributed clouds), and level (i.e., cirrus vs. nimbus clouds).
Secondly, the MODIS cloud detection algorithm incorrectly identifies snow as clouds, in
some situations. Although the MODIS cloud detection algorithm is based on differences
between cloud and snow/ice reflectance and emittance properties at visible and infrared
wavelengths, it cannot always discriminate clouds from snow (Hall et al., 2006).
Third, the MODIS data may be correctly detecting and masking those pixels with highoptically thin cirrus clouds that are not visually apparent on the TM imagery.
It is likely that all three scenarios have a role to play in the differences in cloud cover
observable in the TM data and the MODIS cloud mask. It should be noted that although
cloud cover will impact the accurate determination of the snow fraction in a scene, the
SWE estimates derived from passive microwave satellite imagery are not affected by
clouds because electromagnetic radiation is not attenuated by atmospheric moisture at 18
and 37 GHz (the frequencies used in the MSC SWE algorithm).
109
CHAPTER SIX - CONCLUSIONS
The results presented throughout Chapter Five show that the objectives of this research
have largely been achieved. This section highlights the conclusions of the four main
objectives of this research in the following order:
i. determine how to best represent ground-truth data for a partial snow cover;
ii. determine how well the MSC prairie SWE algorithm performed over a partial
snow cover;
iii. document how spatial variations observed in a partially snow covered area
were manifest in passive microwave estimates;
iv. discover the degree of fractional snow cover that was required to influence a
passive microwave - derived SWE estimate at different spatial resolutions.
6.1
In-situ vs. Passive Microwave SWE Estimates
The first objective was to determine how to best represent ground-truth data for a partial
snow cover. This objective was met by performing z-tests and interpreting the linear
regression outputs following the method proposed by Gupta (2000). Although
statistically significant models were established between many of the 2005 remote
sensing and in-situ data sets, the proportion of the variance in the satellite estimates that
could be explained by the ground observations (i.e., the Model Summaries) was, at most,
0.31. This suggests that the ground data were insufficient for deriving SWE from
spaceborne passive microwave observations and/or that the remote sensing data were
inappropriate. Previous research (reviewed earlier) has shown that it is possible to obtain
reliable SWE estimates through remote sensing where there is a consistent and complete
110
snow cover. The remote sensing data used here differed from that of the previous studies
in that it was coincident with inconsistent and partial snow cover. The continuous snow
cover assumption embedded in the MSC passive microwave SWE algorithm resulted in
an overestimation of SWE. The overestimation of the MSC SWE algorithm for the
remote sensing data sets evaluated, specifically AMSR-E data, confirmed that the
algorithm was inaccurate under patchy and variable snow conditions.
In spite of the poorly articulated regression models, there were several other in-situ
observations that appeared to play an important role in affecting the satellite passive
microwave data. The presence or absence of an ice lens in the snow pack was
consistently identified as a significant coefficient in the regression analyses. Other
observations that may prove to be useful that were not investigated in this research
include the snow depth and the ground temperature. These additional variables in the
regression models indicate that the relationship between passive microwave - derived
SWE estimates and ground measurements over a variable snow cover is potentially nonlinear, and that a multi-variate regression equation may provide more accurate results for
passive microwave SWE estimates across variable snow conditions.
Consideration of patchy snow cover is challenging from a ground sampling perspective,
however the results of the z-tests and outputs of the linear regression models showed that
the actual conditions found at each sampling site should be incorporated in ground-truth
data sets when collecting observations over a partial snow cover.
Ill
6.2
MSC Algorithm Performance over a Partial Snow Cover
The second objective was to determine how well the MSC prairie SWE algorithm
performed over a partial snow cover. The MSC algorithm performed reasonably well on
the SSM/I data sets, but inconsistently on the AMSR-E data sets.
The five SSM/I SWE retrievals all produced results that were within the +/- 20 mm
threshold of the MSC algorithm. However, it was found that these results differed for
SWE retrievals collected from morning and afternoon orbits. While retrievals from the
afternoon orbits were within 5 mm of the ground measurements, retrievals from the
morning orbits slightly overestimated the ground measurements, by 12 mm to 13 mm.
Unlike the SSM/I results, none of the 25 km AMSR-E estimates were found to be within
the accuracy of the MSC algorithm, and all of the SWE retrievals were significantly
overestimated. These SWE retrievals overestimated the ground measurements by an
average of 30.1 mm. The minimum SWE retrieval overestimated the ground
measurements by 24.9 mm, while the maximum retrieval was overestimated by 34.4 mm.
As for the 12.5 km AMSR-E estimates, ten out of fifteen (67%) retrievals were found to
be within +/- 20 mm, while the other five (33%) overestimated the ground measurements
by an average of 26.2 mm. The range of these overestimated SWE estimates was 20.2
mm to 44.4 mm.
6.3
Manifestation of Spatial Variability in SWE Estimates
The third objective of this research was to document how spatial variations observed in a
partially snow covered area were manifest in passive microwave estimates.
112
Table 9 (p. 34) shows that variation in land cover can impart considerable variability on
ground measurements. While there was little difference in snow characteristics
representing stubble fields and pastures, there were significant differences in the
characteristics of fallow fields and shelter belts. The results of the statistical models
indicated that several other variables helped to explain the passive microwave - derived
SWE estimates. These included: snow cover percent, the presence or absence of an ice
lens, snow depth and the ground temperature. These variables are considerably different
depending on the type of land cover.
Results of this research showed that the SWE estimates derived from passive microwave
data can produce more accurate results when they are weighted by the fractional snow
cover that is representative of a remote sensing pixel. However, this is conditional on the
sampling sites being evenly distributed and spatially representative of the pixel. As
results from the 2005 field campaign for 12.5 AMSR-E Pixel 1 (Figure 5.10, p. 47)
showed that unevenly distributed and spatially unrepresentative distribution of ground
measurements produce results that make it impossible to accurately validate the satellitederived SWE with ground-truth measurements. Caution should be exercised for passive
microwave footprints containing few ground sampling measurements, as this can produce
misleading results.
113
6.4
Degree of Fractional Snow Cover Influence on SWE Estimates
The fourth, and final, objective of this research was to discover the degree of fractional
snow cover that was required to influence a passive microwave - derived SWE estimate
at different spatial resolutions. With respect to this objective, the results of this research
are inconsistent between the TM - derived and MODIS Fractional Snow Cover images.
The results from the 25 km SSM/I footprint indicate that the degree of fractional snow
cover required to influence a passive microwave SWE retrieval is less than 59% (Figure
5.6, p. 43). However, this is based on a TM fractional snow cover that re-scaled 30 m
pixels across a 25 km footprint. Results from the MODIS Fractional Snow Cover for the
SSM/I footprint differ from the TM results. The MODIS results indicate that a fractional
snow cover below 80% will influence an SSM/I SWE retrieval (Figures 5.29 and 5.31,
pp. 69 and 71, respectively).
Results of the 25 km AMSR-E footprint are inconsistent with the results of the same
SSM/I footprint. The results differ from the SSM/I results for the TM fractional snow
cover, but are similar for the MODIS Fractional Snow Cover. The TM results indicated
that a fractional snow cover greater than 59% was required to influence a 25 km AMSRE estimate (Figure 5.8, p. 45). However, analyses of the same footprint integrated with
the MODIS data showed that a fractional snow cover below 80% will influence an
AMSR-E SWE retrieval (Figures 5.37, p. 77). Thus, the MODIS results were nearly
identical between the 25 km AMSR-E and SSM/I results.
114
Similar to the results from the 25 km AMSR-E footprint, results of the 12.5 km AMSR-E
footprints were inconsistent. The degree of fractional snow cover required to influence a
SWE retrieval for TM imagery ranged from 39% (Figure 5.18, p. 56) to 89% (Figure
5.14, p. 51). However, fractional snow covers of 52% and 67% (Figures 5.16 and 5.20,
pp. 53 and 58, respectively) did not appear to influence SWE estimates. Although a 92%
fractional snow cover, collected from the 2008 validation campaign (Figure 5.53, p. 93)
confirmed the upper limit, these results are largely inconclusive. This is partly due to: i)
there were few ground sampling sites coincident with each of the 12.5 km AMSR-E
pixels, and therefore, the characteristics of those sites were likely not representative of
the entire 12.5 km footprint, and ii) the TM fractional snow cover re-scaled 30 m pixels
across the 12.5 km footprints. With respect to the MODIS imagery, the degree of
fractional snow cover required to influence a SWE retrieval is again less than 80%
(Figure 5.41, 5.43 and 5.47, pp. 81, 83 and 87, respectively). While the results of Pixel 6
(Figure 5.49, p. 89) show that SWE estimates for an 86% snow cover was influenced and
a 55% snow cover was not, the characteristics of these individual sampling sites may not
have been characteristic of the 500 m MODIS pixel. Therefore, it is again important to
note that caution should be exercised for passive microwave footprints containing few
ground sampling measurements, as this can produce misleading results.
115
6.5
Summary
In summary, the findings of this study support the following conclusions:
1. Actual conditions found at each sampling site should be incorporated in groundtruth data sets when collecting observations over a partial snow cover;
2. Passive microwave derived SWE estimates tend to overestimate ground SWE
measurements over patchy snow covered areas, particularly AMSR-E derived
SWE estimates and data collected from morning orbits;
3. Passive microwave SWE estimates show improved accuracy when they are
weighted by fractional snow cover derived from optical imagery;
4. MODIS Fractional Snow Cover imagery produces similar results to classified
Landsat TM imagery when the passive microwave footprints are classified as
fractional snow;
5. The degree of fractional snow cover required to influence a passive microwave
derived SWE estimate at different spatial resolutions was:
a) inconclusive for re-scaled 30 m TM imagery, although:
a 59% snow cover did not influence the 25 km SSM/I estimate,
a 59% snow cover did influence the 25 km AMSR-E estimate, and
a 92% snow cover did influence a 12.5 km AMSR-E estimate.
b) below 80% for 500 m MODIS Fractional Snow Cover for all 25 km and
12.5 km passive microwave estimates.
116
6.6
Future Research
Future research should investigate the potential of multi-variate regression models to
derive SWE estimates from passive microwave brightness temperatures for areas with
highly variable snow conditions. Numerous variables were considered significant in the
linear regression models analyzed in this study, but these variables are difficult to account
for in linear equations.
Further research should also focus on more micro-scale analyses of fractional snow
cover. In particular, field data campaigns should be designed with denser sampling
strategies to further investigate the impact of patchy snow cover on numerous 500 m
MODIS pixels. Data collected from one sampling site that is representative of a 30 m
area may not be sufficient to represent a fractional snow cover estimate across a 500 m
pixel. It would also prove useful to extend the micro-scale analyses to include higher
resolution optical imagery, such as the Landsat TM imagery used in this study. For
example, the Landsat TM image could be spatially sub-set to correspond with a MODIS
pixel, and the fractional snow cover of the TM image calculated based on the re-sized
image. This would provide a useful comparison between the MODIS and TM images,
and may provide a better understanding of the impacts of fractional snow cover on
passive microwave SWE estimates.
Finally, additional research should incorporate Surface Reflectance MODIS data. The
Surface Reflectance data, transformed using the same iNDSI algorithm as used for the
TM data, would provide a better comparison between the MODIS and TM sensors than
the MODIS Fractional Snow Cover product did, and it would also eliminate the cloud
117
mask. This would also lead to a better understanding of the impacts that a partial snow
cover has on SWE estimates derived from passive microwave brightness temperatures at
different spatial resolutions.
118
REFERENCES
Armstrong, R.L. and M.J. Brodzik, 1995. An Earth-gridded SSM/I data set for
cryospheric studies and global change monitoring, Advances in Space Research
16(10): 155-163.
Brown, Ross D., Anne E. Walker, and Barry Goodison, 2000. "Seasonal Snow Cover
Monitoring in Canada: An Assessment of Canadian Contributions for Global
Climate Monitoring," Proceedings, 57th Eastern Snow Conference, Syracuse,
New York, USA, 2000, obtained November 20th, 2004, from:
http://www.easternsnow.org/proceedings/2000/brown.pdf.
Bussieres, Normand, Danielle de Seve and Anne Walker, 2002. "Evaluation of MODIS
Snow Cover Products Over Canadian Regions," in 2002 IEEE International
Geoscience and Remote Sensing Symposium and the 24l' Canadian Symposium
on Remote Sensing, June 24 - 28, 2002 Toronto, Ontario, Canada, CD-ROM.
Comiso, J. C. and F. Nishio (2008). "Trends in the sea ice cover using enhanced and
compatible AMSR-E, SSM/I, and SMMR data," Journal of Geophysical
Research, 113 (C02S07).
De Abreu, Roger, and Ellsworth LeDrew, 1997. "Monitoring Snow and Ice Conditions
Using a Normalized Difference Index Based On AVHRR Channels 1 and 2,"
obtained July 28, 2003, from:
http://www.crvsvs.uwaterloo.ca/science/documents/ger97 deabreu2.pdf.
Derksen, C , A. Walker, E. LeDrew, B. Goodison, 2002a. "Time-series analysis of
passive-microwave-derived central North American snow water equivalent
imagery," Annals ofGlaciology, Vol. 34, pp. 1-7.
Derksen, C , R. Brown, A. Walker, and B. Brasnett, 2002b. "Comparison of Model,
Snow Course, and Passive Microwave Derived Snow Water Equivalent Data for
Western North America," Proceedings, 59th Eastern Snow Conference, Stowe,
Vermont, USA, 2002, obtained November 20th, 2004, from:
http://www.easternsnow.Org/proceedings/2002/017_Derksen.pdf.
Derksen, C , A. Walker, B. Goodison, 2003a. "A comparison of 18 winter seasons of in
situ and passive microwave-derived snow water equivalent estimates in Western
Canada," Remote Sensing of Environment, 88 (2003) pp. 271-282.
119
Derksen, C , and A. Walker, E. LeDrew, B. Goodison, 2003b. "Combining SMMR and
SSM/I Data for Time Series Analysis of Central North American Snow Water
Equivalent," Journal ofHydrometeorology, Vol. 4, pp. 304-316.
Goita, K., A.E. Walker, B.E. Goodison, and A.T.C. Chang, 1997. "Estimation of Snow
Water Equivalent in the Boreal Forest Using Passive Microwave Data,"
Proceedings, GER '97 (International Symposium: Geomatics in the Ear of
Radarsat), Ottawa, Canada, May 25-30, 1997. CD-ROM published by Natural
Resources Canada and National Defence.
Goita, K., A. Walker, and B. Goodison, 2003. "Algorithm Development for the
Estimation of Snow Water Equivalent in the Boreal Forest using Passive
Microwave Data," International Journal of Remote Sensing. 24(5): 1097-1102.
Goodison, B., 1989. Determination of areal snow water equivalent on the Canadian
Prairies using passive microwave satellite data. Proceedings, 1989 International
Geoscience and Remote Sensing Symposium, Vancouver. 3:1243-1246.
Goodison, B.E., 1995. "Remote Sensing of Snow Cover," in The Canadian Remote
Sensing Contribution To Understanding Global Change, LeDrew, E., M. Strome,
and F. Hegyi (eds.), Department of Geography Publication Series, No. 38,
University of Waterloo.
Goodison, B.E. and A.E. Walker, 1993. "Use of Snow Cover Derived from Satellite
Passive Microwave Data as an Indicator of Climate Change," Annals of
Glaciology, Vol. 17, pp. 137-142.
Goodison, B.E. and A.E. Walker, 1994. "Canadian development and use of snow cover
information from passive microwave satellite data," in Passive Microwave
Remote Sensing of Land-Atmosphere Interactions, Choudhury, B.J., Y.H. Kerr,
E.G. Njoku, and P. Pampaloni (eds.), VSP, Utrecht, The Netherlands, pp. 245262.
Goodison, B.E. and A.E. Walker, 1995. "Canadian development and use of snow cover
information from passive microwave satellite data," in Passive Microwave
Remote Sensing of Land-Atmosphere Interactions, Choudhury, B.J., Y.H. Kerr,
E.G. Njoku, and P. Pampaloni (eds.), VSP, Utrecht, The Netherlands, pp. 245262.
Gupta, V., 2000. Regression explained in simple terms. Obtained online on April 24,
2005, from
http://www.telecom.csuhavward.edu/~esuess/Links/Software/RegressionExplaine
d/regressionexplained.doc
120
Hall, Dorothy K., George A. Riggs, and Vincent V. Salomonson, 2006, updated
daily. MODIS/Terra Snow Cover Daily L3 Global 500m Grid V005, [February
21-23, 2005, and March 2, 2008]. Boulder, Colorado USA: National Snow and
Ice Data Center. Digital media.
Hare, F.K. and M.K. Thomas, 1979. Climate Canada, 2nd Ed., Canada: Wiley.
Imaoka, Keiji, Toshihoro Sezai, Toshiaki Takeshima, Toneo Kawanishi, and Akira
Shibata, 2002. "Instrument characteristics and calibration of AMSR and AMSRE," in 2002 IEEE International Geoscience and Remote Sensing Symposium and
the 24" Canadian Symposium on Remote Sensing, June 24 - 28, 2002 Toronto,
Ontario, Canada, CD-ROM.
Laycock, Arleigh H., 1972. "The Diversity of the Physical Landscape," in The Prairie
Provinces, Smith, P.J. (ed.), University of Toronto Press.
O'Sullivan, David and David Unwin, 2003. Geographic Information Analysis, John
Wiley & Sons, Inc., Hoboken, New Jersey.
Rees, W. Gareth, 2006. Remote Sensing of Snow and Ice. CRC Press, Taylor & Francis
Group, Boca Raton, FL. Ch. 5 (pp. 137-155).
Robinson, David A., 1993. "Monitoring Northern Hemisphere Snow Cover,"
Proceedings, Snow Watch '92: Detection Strategies for Snow and Ice, Niagaraon-the-Lake Canada, March 1993. Waterloo Canada: Earth Observations
Laboratory, Institute for Space and Terrestrial Science, ISTS-EOL-TR93-005, pp.
1-31.
Shibata, Akira, 2002. "AMSR/AMSR-E Data Utilization Plan," in 2002 IEEE
International Geoscience and Remote Sensing Symposium and the 24 Canadian
Symposium on Remote Sensing, June 24 - 28, 2002 Toronto, Ontario, Canada,
CD-ROM.
Singh, Purushottam Raj and Thian Yew Gan, 2000. "Retrieval of Snow Water
Equivalent Using Passive Microwave Brightness Temperature Data," Remote
Sensing of Environment, Vol. 74, pp. 275-286.
Sokol, J., T.J. Pultz, and A.E.Walker, 1999. "Passive and Active Airborne Microwave
Remote Sensing of Snow Cover," Proceedings, 4th International Airborne Remote
Sensing Conference/21st Canadian Symposium on Remote Sensing.
121
Thraves, Bernard D., M.L. Lewry, Janis E. Dale and Hansgeorg Schlichtmann (eds.),
2007. Saskatchewan: Geographic Perspectives. Canadian Plains Research Center,
University of Regina, Regina, Saskatchewan, CA. Ch. 2 (pp. 28-31).
Turchenek, K., 2004. "Validation of Snow Water Equivalents Derived from Passive
Microwave Brightness Temperatures over a Prairie Environment," Unpublished
Undergraduate Thesis, submitted to, and accepted by, the Department of
Geography at the University of Regina, Saskatchewan.
Walker, A.E. and B.E. Goodison, 1993. "Discrimination of a Wet Snow Cover Using
Passive Microwave Satellite Data," Annals ofGlaciology, Vol. 17, pp. 307-311.
Walker, A. and B. Goodison, 2000. "Challenges in Determining Snow Water Equivalent
over Canada using Microwave Radiometry," Proceedings, International
Geoscience and Remote Sensing Symposium, Honolulu, July, 2000, pp. 15511554.
Walker, Anne, Barry Goodison, Michael Davey and David Olson, 1995. Atlas of
Southern Canadian Prairies Winter Snow Cover from Satellite Passive Microwave
Data: November 1978 to March 1986, Downsview, Canada: Atmospheric
Environment Service, Environment Canada.
Walker, A.E, J.W. Strapp, and J.I. MacPherson, 2002. "A Canadian Twin Otter
Microwave Radiometer Installation for Airborne Remote Sensing of Snow, Ice
and Soil Moisture," in 2002 IEEE International Geoscience and Remote Sensing
Symposium and the 24* Canadian Symposium on Remote Sensing, June 24 - 28,
2002 Toronto, Ontario, Canada, CD-ROM.
Winther, Jan-Gunnar, 1993. "Studies of Snow Surface Characteristics by Landsat TM in
Dronning Maud Land, Antarctica," Annals ofGlaciology, Vol. 17, pp. 27-34.
122
Документ
Категория
Без категории
Просмотров
0
Размер файла
5 573 Кб
Теги
sdewsdweddes
1/--страниц
Пожаловаться на содержимое документа