close

Вход

Забыли?

вход по аккаунту

?

j.scitotenv.2017.09.335

код для вставкиСкачать
STOTEN-24223; No of Pages 12
Science of the Total Environment xxx (2017) xxx–xxx
Contents lists available at ScienceDirect
Science of the Total Environment
journal homepage: www.elsevier.com/locate/scitotenv
Utility of ASTER and Landsat for quantifying hydrochemical
concentrations in abandoned gold mining
Solomon G. Tesfamichael ⁎, Aros Ndlovu
Department of Geography, Environmental Management and Energy Studies, University of Johannesburg, South Africa
H I G H L I G H T S
G R A P H I C A L
A B S T R A C T
• Remote sensing was used to quantify several gold mining related hydrochemicals.
• ASTER and Landsat data served as predictors in an information theoretic approach.
• Both data showed promising estimation
capability for most hydrochemicals.
• Trace elements generally had low correlation with spectra.
• Improved sampling is encouraged for
better estimation accuracy.
a r t i c l e
i n f o
Article history:
Received 11 July 2017
Received in revised form 29 September 2017
Accepted 30 September 2017
Available online xxxx
Editor: F.M. Tack
Keywords:
Mine contaminated water
Hydrochemical properties
Landsat
ASTER
Remote sensing
Akaike's Information Criterion
a b s t r a c t
The effect of mining on water resources is severe and requires careful monitoring and management. Remote
sensing has been used to characterize water quality indicators in efforts to fight mine-induced contamination.
Much focus has however been placed on producing a qualitative classification of water qualities. Moreover, the
number of variables considered in most studies is relatively small for a large number of hydrochemical constituents common in water bodies associated with gold mining activities. This study is aimed at quantifying a comprehensive list of field- and laboratory-measured chemical constituents of water samples from abandoned mines
using remotely-sensed data. Akaike's Information Criterion was used to estimate each of the constituents using
statistical values derived from individual bands of ASTER and Landsat data as predictors. Fairly good accuracies
were obtained for constituents such as redox potential (Eh), major anions and cations. In contrast, trace elements
correlated poorly with ASTER and Landsat bands, due mainly to a sampling anomaly. The performances of the
two images in estimating the constituents were comparable. These findings suggest the potential of multispectral, moderate spatial resolution remote sensing for quantifying different hydrochemical properties of water bodies in mining environments. Further studies are however encouraged to enhance accuracies and reliability using a
greater number of samples than was used in this study to capture the variability present in the population.
© 2017 Elsevier B.V. All rights reserved.
1. Introduction
Mining has a significant contribution to South Africa's economy and
livelihoods; however, it is associated with various negative impacts
(McCarthy, 2011). One such impact is acid mine drainage (AMD)
⁎ Corresponding author.
E-mail address: sgtesfamichael@uj.ac.za (S.G. Tesfamichael).
which refers to water pollution due to reaction with minerals under favorable conditions. For example, pyrite and sulphates are stable minerals underground, but release acidic solutions - iron hydroxide and
sulphuric acid - when reacting with oxygenated water (Scott, 1995;
Naicker et al., 2003; Johnson and Hallberg, 2005). These solutions dissolve ore minerals from which heavy metals such as manganese, aluminium and iron are released into the water. This problem is common
in mine dumps where proper mine closure and reclamation procedures
https://doi.org/10.1016/j.scitotenv.2017.09.335
0048-9697/© 2017 Elsevier B.V. All rights reserved.
Please cite this article as: Tesfamichael, S.G., Ndlovu, A., Utility of ASTER and Landsat for quantifying hydrochemical concentrations in abandoned
gold mining, Sci Total Environ (2017), https://doi.org/10.1016/j.scitotenv.2017.09.335
2
S.G. Tesfamichael, A. Ndlovu / Science of the Total Environment xxx (2017) xxx–xxx
did not take place (McCarthy, 2011; Durand, 2012; Khalil et al., 2014). If
unattended, acid mine drainage has adverse effects on rivers, underground water systems, aquatic life, animal husbandry (e.g. McIntyre
et al., 2016) and crop productivity (Choudhury et al., 2017).
Acquiring accurate and timely information on the level of water pollution due to mining related activities is an important step in managing
the problem (Shang et al., 2009; Yenilmez et al., 2011). Traditionally, the
effects of mine-induced environmental hazard are monitored by
collecting data using field surveys and subsequently analyzing them in
the laboratory. Such methods are time consuming, expensive and are
ineffective in capturing the spatial variability of pollution, particularly
in large areas. Remote sensing, on the other hand, has a great potential
in complementing field-based surveys for assessing mining related environmental variables. Basically, remote sensing uses reflectance or absorbance properties of electromagnetic energy to characterize features.
Key advantages of the technology include wide spatial coverage, availability of spectral information acquired in different electromagnetic radiations and periodic data acquisition modes that allow for water
quality monitoring purposes.
The performance of remote sensing in quantifying water qualities
particularly in large and natural water bodies is well documented
(e.g., Wang et al., 2006; Zhao et al., 2011; Bonansea et al., 2015; Dube
et al., 2015; Urbanski et al., 2016). Water quality indicators in such
water bodies are usually summarized into generic properties such as
suspended solids, Secchi disk depth and turbidity (e.g. Wang et al.,
2006), albeit more variables are considered occasionally (e.g. Torbick
et al., 2013). Water bodies associated with mining activities harbor relatively large number of minerals (Lottermoser, 2003). Moreover, they
cover small spatial areas since they are mostly developed to support
mining operations or are water accumulated in abandoned mine pits.
These problems necessitate for the use of remote sensing systems
with better resolutions than those used in natural and large water bodies. Sams and Veloski (2003), for example, utilized a high spatial resolution (1 m2) night-time thermal infrared imagery to identify acid mine
drainage source sites in a coal mining area in Pennsylvania, and reported a 53% accuracy of identifying the sites. While the high spatial resolution was advantageous to spatially characterize water quality in a great
deal of locational detail, such data are often commercial and are unavailable in the public domain. Yucel et al. (2014) used multi-temporal and
multi-source remotely-sensed (Landsat, Quickbird and Worldview)
data to assess the areal dynamics of acid mine lakes in northwest
Turkey, and reported an increase in the lakes. Although the study quantified hydrochemical properties including physical parameters, anions,
cations and trace elements, it did not relate them to the spectra of
remotely-sensed data. Furthermore, it is unknown whether the single
sample used per lake was representative of the variability present in
the hydrochemical properties within each lake. Davies and Calvin
(2016) profiled spectra against iron concentration in an acid mining
drainage system in a controlled laboratory set-up using spectral measurements of synthetic data and samples collected from actual water
bodies. Although such an approach is an important initial step in
informing the potential of remote sensing for feature characterization,
it does not fully mimic actual scenarios of airborne or spaceborne systems that are influenced by atmospheric interferences. Moreover, the
study focused on characterization of acid mine drainage using spectral
curve profiles, and did not relate chemical concentrations and spectra
quantitatively.
There is a need to extend the utility of remote sensing to quantify
physicochemical concentrations of water bodies in mining environments. Research efforts in this regard are rare. Schroeter and Gläβer
(2011), for example, applied Spearman–Rho correlation analysis to
identify the existence of linear relationship between limited number
of hydrochemicals (pH, Secchi depth, chlorophyll-a, dissolved organic
carbon, dissolved iron and total phosphorus) and reflectance values derived from Landsat's visible bands (red, green and blue). The correlation
was however used as a precursor for clustering analysis instead of
quantifying concentrations. Ong and Cudahy (2014) developed a predictive model to quantify pH in an abandoned pyrite mine in Australia
using airborne-based hyperspectral imagery. Apart from focusing on
pH only, the study was conducted on soils rather than on water bodies.
Such limited or indirect quantitative analyses do not fully address the
complex characteristics of mine-related water bodies (Lottermoser,
2003; Kopačková, 2014).
This study is aimed at exploring the potential of multispectral remote sensing in quantitative modelling of several geochemical concentrations in mine-affected water bodies. A secondary aim of the study
was to compare the performances of Landsat and ASTER images in
quantifying geochemical contents. Although there are certain similarities in spectral characteristics of ASTER and Landsat data, the former
had better spectral resolution in the shortwave infrared region while
it lacks the blue wavelength. It is unknown how such differences influence quantification of hydrochemical concentrations. A statistical approach that provides alternative estimation models was adopted in
the study. The approach was preferred due to the lack of established relationships in previous studies as well as the high level of variability in
hydrochemical values used in the study. As such, the study can be considered an exploratory analysis that provides an insight into the potential of quantifying hydrochemicals using a cheap, rapid and moderately
detailed remote sensing based approach. It is however important to account for local conditions when considering model transfer due to the
high level of variability in mining environments (Kopačková, 2014). A
notable limitation of this study is the relatively small sample size that
could not capture the variability of many of the hydrochemicals in sufficient number; undertaking similar studies is therefore vital to increase
the database from which more robust models can be developed.
2. Methodology
2.1. Study area
The study by Tutu et al. (2008), which our study extracted reference
data from, was conducted in the Witwatersrand goldfields (Fig. 1). The
region is known for intensive gold mining activities that has been taking
place for more than a century (Handley, 2004) and a source of livelihoods for the City of Johannesburg (Durand, 2012). Climatically, the region receives mean annual rainfall of approximately 604 mm, most of
which occurs from late spring to early autumn (October – March),
while the temperature varies between 0.8 °C in winter to 26 °C in summer. The area has a network of rivers and wetlands most of which drain
into the Klip River, which in turn is a tributary to one of South Africa's
major rivers – the Vaal River. Both the Klip River and the Vaal River supply major ecosystem services for domestic, agricultural, industrial and
recreational purposes for several areas, including the City of Johannesburg (McCarthy et al., 2007). Overall, water pollution due to mining
within the drainage system decreases with distance from mining sites;
however this is primarily as a result of the nearby wetlands serving as
a sink (Davidson, 2003; McCarthy et al., 2007).
2.2. Reference data
Reference data used for this study was obtained from Tutu et al.
(2008), who quantified various chemical characteristics through analysis of surface and ground water samples collected from a number of
water bodies. All the water bodies were located in the vicinity of abandoned gold mine sites (Tutu et al., 2008). The samples were collected in
April 2002 and August 2003, representing the end of wet and dry seasons, respectively. Field parameters used in the study included pH,
redox potential (Eh) and electrical conductivity (EC) while chemical
characteristics included Sulphate (SO₄2 −), Nitrate (NO₃−), Chlorine
(CI−), Sodium (Na), Potassium (K), Calcium (Ca), Magnesium (Mg), Aluminium (Al), Copper (Cu), Cobalt (Co), Iron (Fe), Manganese (Mn),
Nickel (Ni), Zinc (Zn) and Uranium (U). The geographical coordinates
Please cite this article as: Tesfamichael, S.G., Ndlovu, A., Utility of ASTER and Landsat for quantifying hydrochemical concentrations in abandoned
gold mining, Sci Total Environ (2017), https://doi.org/10.1016/j.scitotenv.2017.09.335
S.G. Tesfamichael, A. Ndlovu / Science of the Total Environment xxx (2017) xxx–xxx
3
Fig. 1. Study area located in Gauteng province, South Africa. (A) shows a Google Earth™ image outlining the Gold mine area in which gold-coloured land covers represent open-pit gold
mines, while field sample points are shown spread within the area. (A)–(F) show Landsat (left) and ASTER (right) images overlaid with the digitized zones from which statistical predictors
were computed. The zones illustrate sampling from dams ((A) and (B)), wide streams ((C) and (D)) and narrow streams ((E) and (F)).
of each sampling point was recorded (Tutu et al., 2008). These points
were plotted within the GIS environment and subsequently overlaid
on Google Earth™. The overlay showed that certain water bodies had
two or more sampling points with a minimum in-between distance of
~ 450 m; these samples were treated independently to avoid any assumption that all chemicals are distributed uniformly within a water
body. Besides, the distances between points that are within a water
body were relatively large and thus aggregating such samples would involve suppressing variations of several pixels of Landsat and ASTER images used in the study. It should also be noted that Tutu et al. (2008)
treated each sample independently of the others. The surface water
bodies in April 2002 included 30 streams, 14 dams and 3 wetlands,
while the August 2003 dataset included 24 samples from streams and
10 from dams but none from the wetlands. A further screening of
these samples was carried out to determine the final sample set, as described in the next section.
2.3. Digitizing sampling zones
A common problem associated with data acquisition using remote
sensing is the effect on reflectance of path radiance in the atmosphere.
The path of electromagnetic radiation reflecting from the target can be
distorted by atmospheric components, resulting in misplaced pixels.
This can be mitigated by taking an average of multiple pixels found
within a neighbourhood referred to here as sampling zones, which
were delineated on Landsat image while Google Earth™ was used for
validation where temporally coincident images were available.
Conservative rules were followed in the delineation process. Firstly,
only samples that fell within pixels that were interpreted as water bodies were accepted as reliable samples. Secondly, delineation of pixels
near or around an acceptable sample was limited to pixels within a relatively short distance to maintain the spatially localized nature of any of
the chemicals under investigation. It should be noted that this conforms
to the original study (Tutu et al., 2008) which aimed at assessing the
parameters and chemicals at a localized scale and thus collected up
to three samples per water body. Thirdly, the delineation included
mostly pixels that were considered to be water bodies within the
neighbourhood (Fig. 1A–F), while pixels that appeared representing
vegetation were avoided. Fourthly, the shape of a neighbourhood was
dictated by the shape of a water body; accordingly, neighbourhoods of
pixels in dams and in wide streams were somewhat round (Fig. 1C
and D) while they had linear shapes along narrow streams (Fig. 1E
and F) to avoid incorporating non-water pixels. Enforcing these rules resulted in the elimination of all samples from the wetlands, 12 from
streams and one from a dam due to significant vegetation cover on or
around the samples from the April 2002 data. The wetlands in particular
were dominated by Phragmites and Typha spp. reeds as well as peats
(Tutu et al., 2008). Consequently, a total of 31 samples were deemed
fit for further analysis involving remotely-sensed data. Similarly, nine
samples from streams and one sample from a dam were eliminated
from the August 2003 data resulting in a total of 25 usable samples.
Few of the streams from which some of the stream and dam samples
were taken are classified as first order rivers (FEPA, 2011), while others
did not have class designation. The widths of the streams surrounding
the final delineated zones varied between ~ 2–5 pixels (1 pixel =
30 m). The large widths in some instances were due to the streams
draining from paddocks or ponds near dumps and tailings. Samples
from streams were collected in the thalwegs (Tutu et al., 2008),
allowing for delineating vegetation-free pixels.
The locations of the August 2003 samples coincide with the April
2002 samples. Descriptive statistics of chemicals computed for these
samples are presented in Table 1. pH and Eh have lower variations
than the other field parameters and chemical concentrations observed
in the laboratory in both datasets, as can be seen from coefficients of
Please cite this article as: Tesfamichael, S.G., Ndlovu, A., Utility of ASTER and Landsat for quantifying hydrochemical concentrations in abandoned
gold mining, Sci Total Environ (2017), https://doi.org/10.1016/j.scitotenv.2017.09.335
4
S.G. Tesfamichael, A. Ndlovu / Science of the Total Environment xxx (2017) xxx–xxx
Table 1
Water quality variables obtained from field survey and laboratory analyses of water samples (n = 31) and used in the study. Values in brackets represent the August 2013 dataset (n = 25).
Source: Tutu et al. (2008).
Variables
Field parameters
Major anions
Major cations
Trace elements
pH
Eh (mV)
EC (mS cm−1)
−1
SO2−
)
4 (mg L
−1
NO−
)
3 (mg L
−
−1
Cl (mg L )
Na (mg L−1)
K (mg L−1)
Ca (mg L−1)
Mg (mg L−1)
Al (mg L−1)
Cu (mg L−1)
Co (mg L−1)
Fe (mg L−1)
Mn (mg L−1)
Ni (mg L−1)
Zn (mg L−1)
U (mg L−1)
Minimum
Maximum
Mean
Median
SD
Coefficient of variation (%)
3 (4)
311 (368)
b1 (b1)
13 (16)
b1 (b1)
b1 (13)
8 (8)
3 (4)
25 (24)
7 (9)
b1 (b1)
b1 (b1)
b1 (b1)
b1 (b1)
b1 (b1)
b1 (b1)
b1 (b1)
b1 (b1)
9 (9)
711 (670)
7 (b1)
5080 (3099)
11 (83)
98 (136)
195 (176)
80 (56)
202 (679)
94 (142)
431 (45)
6 (3)
38 (6)
270 (116)
119 (55)
71 (8)
108 (4)
24 (20)
6 (7)
545 (505)
1 (b1)
806 (788)
5 (19)
39 (57)
50 (38)
14 (17)
83 (206)
33 (38)
19 (4)
1 (b1)
2 (1)
23 (9)
13 (9)
4 (1)
4 (b1)
2 (1)
7 (7)
580 (467)
1 (b1)
337 (353)
5 (3)
30 (51)
36 (24)
9 (14)
74 (51)
36 (19)
b1 (b1)
b1 (b1)
b1 (b1)
1 (b1)
3 (1)
b1 (b1)
b1 (b1)
b1 (b1)
2 (1)
116 (90)
1 (b1)
1228 (968)
3 (29)
28 (38)
46 (42)
18 (13)
50 (236)
21 (39)
77 (10)
1 (b1)
7 (1)
59 (26)
25 (17)
13 (b2)
19 (1)
4 (4)
33 (18)
21 (18)
106 (89)
152 (123)
56 (155)
71 (67)
91 (109)
125 (77)
61 (115)
63 (105)
396 (272)
150 (349)
337 (212)
256 (293)
198 (193)
343 (205)
467 (196)
271 (458)
variation (CV). In fact, most of the variables show a great deal of variation, with all trace elements having a CV exceeding 100%. Considerable
differences in the mean and median values of the variables are indications of the lack of normal distribution in the data; instead, it indicates
isolated or few samples falling in the lower end of the data distribution.
However, observation of these isolated values was not consistent for the
two dates. For example, the maximum NO−
3 concentrations for April
2002 and August 2003, respectively, were 11 and 83 mg L−1, resulting
in a larger CV in the latter case.
2.4. Remotely sensed data
Remotely sensed images used for the study included Landsat-7 and
Advanced Spaceborne Thermal Emission and Reflection Radiometer
(ASTER) images. Both images were acquired from the United States
Geological Survey (USGS) online portal (www.earthexplorer.usgs.
gov). Images acquired on the dates closer to April 2002 (date of the reference data) had significant cloud covers, and thus were unsuitable for
the study. Although atmospheric correction is an option to remove
clouds, it is often associated with uncertainties and may even introduce
errors to the pixel values. As a result, Landsat-7 image (path: 170 and
row: 78) acquired on the 28th of March 2002 and ASTER image acquired
on the 7th of January 2002 were utilized for the study. These dates were
deemed acceptable, since they more or less fall within the same season
(summer or end of summer) as the field survey time. Moreover, no variation was expected in the amount of chemicals in the water bodies considering that there were no mining activities that could induce
significant change between the field and remote sensing survey times.
Landsat image temporally closest to the August 2003 field data was acquired on the 18th November 2003; this date was considered to be in
the same season as the field survey. In contrast, the only ASTER images
covering all samples were acquired in March 2003 and in April 2004.
Both dates were outside of the field survey season and therefore could
not be used in the study.
Table 2 presents summaries of the images used in the study. ASTER
image had more spectral bands than Landsat did in the visible-toinfrared regions (9 vs. 6 or 7), while it had a better spatial resolution
in the same electromagnetic energy regions. Most of the spectral differences between Landsat and ASTER images is in the shortwave-infrared
region, in which ASTER has more resolution. Thermal bands for both
ASTER and Landsat imaging systems have coarser spatial resolutions
relative to the sampling units of this study, and therefore were not
used in the study.
Pixel values of Landsat and ASTER images are delivered by data providers as digital numbers (DNs) that are integers scaled to reduce the
size of an image file while preserving the relative differences among
pixels. It is therefore necessary to scale these values back to the original
reflectance if quantitative analysis relating reflectance and chemical
properties such as one aimed in this study is sought. This is achieved
through radiometric calibration of individual bands by converting DN
to reflectance value of each pixel using rescaling specifications provided
with the data (Chander et al., 2009). The resultant reflectance is a measure of the proportion of reflected radiation relative to that incident on
the target, and thus is a unitless value. In this study, top-of-atmosphere
reflectance was used as the final dataset since atmospheric correction
was not applied to avoid introduction of error that cannot be quantified.
Table 2
Band specifications of Landsat-7 and ASTER images used to correlate with April 2002 field data, and Landsat-5 TM used correlate with August 2003 field data. Band names given in brackets
for ASTER are to show similarity with Landsat-5 and Landsat-7 bands. Note that Landsat-5 and landsat-7 have similar band designations and spatial resolutions, except for lack of a panchromatic band in the former case.
Year 2002 imagery
Year 2003 imagery
Landsat-7 ETM+
ASTER
Landsat 5-TM
Band
Wavelength width (μm)
Spatial resolution (m)
Band
Wavelength width (μm)
Spatial resolution (m)
Blue
Green
Red
NIR
SWIR 1
SWIR 2
Panchromatic
–
–
0.45–0.52
0.52–0.601
0.63–0.69
0.77–0.90
1.55–1.75
2.06–2.35
0.52–0.90
–
–
30
30
30
30
30
30
15
–
–
VNIR (Green)
VNIR (Red)
VNIR (NIR)
SWIR 1
SWIR 2
SWIR 3
SWIR 4
SWIR 5
SWIR 6
0.52–0.60
0.63–0.69
0.78–0.86
1.60–1.70
2.15–2.19
2.19–2.23
2.24–2.29
2.30–2.4
2.36–2.43
15
15
15
30
30
30
30
30
30
0.45–0.52
0.52–0.60
0.63–0.69
0.76–0.90
1.55–1.75
2.08–2.35
–
–
–
Please cite this article as: Tesfamichael, S.G., Ndlovu, A., Utility of ASTER and Landsat for quantifying hydrochemical concentrations in abandoned
gold mining, Sci Total Environ (2017), https://doi.org/10.1016/j.scitotenv.2017.09.335
S.G. Tesfamichael, A. Ndlovu / Science of the Total Environment xxx (2017) xxx–xxx
Zonal statistics were extracted from reflectance values of each band
using boundaries digitized around field reference points as zones. The
extraction was implemented in ArcGIS (ESRI®ArcGIS, version 10.5,
380 New York Street, Redlands, CA). The statistics included minimum,
maximum, mean, range, standard deviation and sum of reflectance of
pixels per zone. Statistics such as variance was not included in this
study since it is strongly related to standard deviation. It is unclear as
to how these statistics, individually or in combination, relate to chemical
characteristics since no study attempted this approach previously. The
choice of these statistics in this study is an attempt to explore and compare the performances of as many remote sensing based statistics as
possible in quantifying mine-related water pollutions. These bandspecific statistics served as input (independent) variables to estimate
field- and laboratory measured water quality parameters using statistical analysis explained in Section 2.5.
2.5. Statistical analysis
An information-theoretic approach was used to estimate each of the
dependent (field- and laboratory measured parameters) variables using
predicting variables (zonal statistics derived from remotely-sensed
data) in this study. Specifically, Akaike's Information Criterion (AIC)
(Akaike, 1973; Burnham and Anderson, 2002) that builds competing
linear regression models using all predicting variables as individuals as
well as all possible combinations was applied here. Building multiple
models per dependent variable is possible in AIC modelling environment, because the approach does not rely on testing the significance
of models or their predicting components. Instead, it builds all possible
models from which the user can choose based on a priori knowledge
(information) about the data. The approach is therefore ideal for this
study which seeks to explore the potential of remote sensing in quantifying chemical properties outside of controlled laboratory environments. In contrast, other parametric approaches (e.g. stepwise
regression) and non-parametric methods (e.g. decision tree based
methods) which return a single model that satisfies a pre-determined
criterion such as probability or significance levels are unlikely to succeed in observational environments where sources of uncertainties
are many and cannot be controlled fully. The approach works on the
premise that, there exists a model that best represents the true relationship between dependent and independent variables; all other competing models are evaluated based on their comparability with the best
(true) model. This evidently implies that there is loss of information
when attempting to simulate the best model. Such information loss is
quantified using Kullback–Leibler Information or distance (Kullback
and Leibler, 1951), according to which a model that has the lowest distance (information loss) is considered the best in approaching the ideal
model.
One of the tools used to evaluate the goodness of models is the AIC
statistic which is quantified for each model using Eq. (1). This statistic
does not provide absolute physical information, but is used to compare
the relative strength of models; the model with the smallest AIC is considered the best while all other models are measured against the best
based on their distance in AIC units (Burnham and Anderson, 2002).
AIC ¼ −2 log L ^θ þ 2K
ð1Þ
5
1986; Burnham and Anderson, 2002). In the case of a small sample
size, a second-order AIC (AICc) which corrects bias due to overfitting is
advised in which the term 2 K is multiplied by a correction factor as
shown in Eq. (2), where n represents sample size (Hurvich and Tsai,
1989). As a general guide, Burnham and Anderson (2002) recommend
the use of AICc if the n/K ratio is b 40. In this study, the ratios (31/6 for
2002 data and 25/6 for 2003 data) are well within the recommended
limit, and thus Eq. (2) was implemented.
n
AIC c ¼ −2 log L ^θ þ 2K
n−K−1
ð2Þ
The total number of models built using the approach to estimate a
dependent variable is equivalent to 2n-1 from n number of predictors.
Substituting 6 (number of zonal statistics predictors) for n in this
study results in a total of 26-1 = 63 competing models for each of the
dependent variables (field and laboratory observed variables). The
best model that relates the dependent and independent variables is assumed to be one with the lowest AICc value. A model that is within 2
AICc units of the best model (for a given dataset) is regarded as highly
comparable to the best model, while the similarity decreases as
the distance increases (Burnham and Anderson, 2002). Comparing
models is particularly important when the predicting variables are
many, and to minimize model overfitting by using only non-collinear
variables. In this study, for instance, there were correlations (similarities) between reflectance of mean and minimum (r ≥ 70%), mean and
maximum (r ≥ 70%) as well as range and standard deviation zonal statistics (r ≥ 90%). Thus, the use of such collinear variables in a model is justified only if the model has considerable superiority to the closest
comparable model that does not include collinear variables. In this
study, models within 10 AICc units were compared, following the recommendation of Burnham and Anderson (2002) who also consider
such a limit as rather flexible. In addition, root-mean-square-error
(RMSE) and the coefficient of determination (R2) were used in the comparison of models.
The aforementioned procedure was implemented by correlating the
hydrochemical data with each of the images (Landsat-5, Landsat-7 and
ASTER images). A comparison between Landsat based estimations between the 2002 and 2003 data was made only as a crude exploratory exercise, due to the considerable differences between the hydrochemical
values of the two dates (Table 1) that would influence model properties.
The differences between the two datasets can be indicative of the difficulty in attaining optimal sampling that would represent the variability
present in the population. This, in fact, justifies the use of AIC modelling
approach that encourages model development and tuning according to
the sample characteristics. Similarly, Landsat-7 and ASTER bands are
different sources of predictors and thus are not expected to yield similar
models. Consequently, resultant models in this study were compared
using similarities in terms of diagnostic statistics (R2 and RMSE) and
the selected predictors in the final models rather than model transferability that assumes invariant model parameters. The applicability of
models developed in such an approach to unknown (or future) samples
can be explored using large sample sizes that include the range of potential values as well as seasonal factors.
3. Results
where Lð^
θÞ is the maximum likelihood/probability of estimated parameters (^θ) given data and model. ^θ quantifies the effects of explanatory
3.1. Performance of Landsat-5 and Landsat-7 bands
variables on a model and includes the intercept, the regression coefficients and the residual variance. A ‘given model’ refers to the form or
structure of a model with undetermined parameters. It should be
noted that a number of such models are considered from which model
selection proceeds. K is the number of free parameters in the model.
This model in its basic form performs well if the sample size relative
to the number of estimated parameters is large (Sakamoto et al.,
Characteristics of the model that yielded the best estimation potential for each dependent variable are presented here (Table 3). All the
models showed a fairly good comparability with the AICc differences
(ΔAICc) of b 10 units in all cases. It should be noted that a model that integrates at least two predicting zonal statistics was preferred, since such
a model had a better reliability than relying on a single predictor that
can influence the result in a random fashion. If the best model (that
Please cite this article as: Tesfamichael, S.G., Ndlovu, A., Utility of ASTER and Landsat for quantifying hydrochemical concentrations in abandoned
gold mining, Sci Total Environ (2017), https://doi.org/10.1016/j.scitotenv.2017.09.335
6
S.G. Tesfamichael, A. Ndlovu / Science of the Total Environment xxx (2017) xxx–xxx
Table 3
A summary of AICc regression analysis results to estimate chemical variables using zonal statistics derived from individual bands of Landsat-7 acquired in March 2002. ΔAICc refers to the
differences between the model selected in this study and the best model with the lowest AICc value per band. RMSE (%) is the relative RMSE quantified as the ratio of absolute RMSE and the
observed mean chemical value multiplied by 100. Highlighted results show good accuracies based on RMSE followed by R2.
returned the lowest AICc) had two or more predictors, then it was taken
as the final model while its ΔAICc would be zero. In most cases, maximum reflectance, mean reflectance or a combination of the two within
each sampling zone formed part of the best model.
Comparison of estimation accuracies across bands shows considerable differences in R2 and RMSE statistics (Table 3). There were more
consistencies in RMSE than in R2 statistic across bands for each variable.
For example, R2 of pH varied between 0.07 (NIR) and 0.24 (SWIR 1),
while RMSE of the same variable ranged from 32 to 34%. Similar characteristics are noted for most of the variables, including field parameters,
major anions and major cations. However, estimations of trace elements
resulted in RMSE that generally had wider ranges than in the other variable groups. Cobalt (Co), for example, had R2 that varied from 0.47 to
0.76 while RMSE ranged between 179 and 263%. Nevertheless, there is
correlation in the range between R2 and RMSE statistics; this can for instance be seen in Cu and Mn that had the lowest R2 of trace elements,
and returned relatively narrower ranges of both R2 and RMSE statistics.
Generally, the best models for most trace elements had relatively high
R2, compared to most of the models for estimating field parameters,
major anions and major cations. In contrast, the latter two groups performed better in terms of RMSE, although variables such as EC and Na
had fairly high RMSE values across bands.
Fig. 2 illustrates best estimated selected field and chemical variables.
Nitrate (NO−
3 ) had the best correlation between estimated and observed values of all, confirming the high R2 (0.75) and low RMSE
(32%) values. There was fairly good fitness between observed and estimated values for redox potential (Eh) and Calcium (Ca), too; this is in
agreement with the relatively low RMSE values. It is nonetheless important to note the difference in RMSE values of Eh and Ca (20 and 56%, respectively), despite the comparability in R2 (0.25 vs. 0.27). The
difference in RMSE is the result of a poor estimation of a sample that
magnified the overall error of estimation. Unlike the above three variables, the relationship between observed and estimated values of Co
was generally poor; this observation corresponds with the poor relative
RMSE (179%), albeit R2 is high (0.76). A similar weak correlation between estimated and observed values was noted for each of the other
trace elements. A sample with significantly high value was recorded in
all trace elements. However, this sample had values that fell within
the range of each of the other chemical characteristics; therefore excluding the sample as an outlier was not considered as an option.
Another notable observation from Fig. 2 is under- and over-estimation of variables at low and high values, respectively. This observation
was particularly pronounced for Eh and Ca. Estimated Eh values ranged
from 400 to 700 mV, compared to observed values of 300 to 700 mV. Estimated Ca values on the other hand ranged from ~20 to ~120 mg L−1
(excluding the isolated sample), compared to the observed range of
~ 20 to ~ 200 mg L− 1. In contrast, NO−
3 was underestimated the least
with nearly all estimated values falling between 0 and 10 mg L−1, compared to the observed 0 and 12 mg L−1. The levels of underestimations
conform to the coefficients of determination; that is, the greater the R2,
Please cite this article as: Tesfamichael, S.G., Ndlovu, A., Utility of ASTER and Landsat for quantifying hydrochemical concentrations in abandoned
gold mining, Sci Total Environ (2017), https://doi.org/10.1016/j.scitotenv.2017.09.335
S.G. Tesfamichael, A. Ndlovu / Science of the Total Environment xxx (2017) xxx–xxx
7
Fig. 2. Comparison of observed and estimated values using Landsat bands: Redox potential (Eh) was estimated using spectral reflectance of blue band; Nitrate (NO−
3 ) was estimated using
spectral reflectance of red band; Calcium (Ca) and Cobalt (Co) were estimated using spectral reflectance of short-wave infrared band. Dotted line in each figure shows the ideal one-to-one
relationship.
the less the underestimation. Ni is an exception which showed rather
poor correlation between observed and estimated values; however
the fitness line still shows an overall underestimation. Despite these underestimations, there was no serious heteroscedasticity phenomenon
particularly for NO−
3 , Eh and Ca, since the deviations of scatterplots
from the fit line in each case did not vary across the observed values
range. Furthermore, the scatterplots were distributed almost uniformly
above and below the fit line in each case, indicating the consistency of
the models.
Correlations of hydrochemicals acquired in August 2003 and reflectance of Landsat-5 bands acquired in November 2003 are presented in
Table 4. pH, Cl−1, K and U had the best accuracy of field parameters,
major cations, major anions and trace elements, respectively. Comparisons of estimation accuracies between 2002 dataset and 2003 dataset in
terms of RMSE showed mixed results, and corresponded well with coefficients of variation. For example, all field parameters had lower coefficients of variation and were better estimated using the 2003 dataset,
while major anions and cations generally had better accuracies in
2002. Although Eh was not the best estimated parameter, it was still estimated at relatively high accuracy making the estimation comparable
with that obtained using the 2002 dataset. Similar to Landsat-7 based
results, trace elements had relatively low accuracies despite high R2
values. The best estimation accuracies were obtained mostly using
short wavelength bands (visible electromagnetic energy); however longer wavelength bands also returned good accuracies. Most of the best
models included maximum and mean zonal statistics, similar to
Landsat-7 based models.
3.2. Performance of ASTER bands
Estimations using ASTER bands were mostly similar to those obtained based on Landsat-7 bands. Green, red, NIR and SWIR 1 bands in
Landsat-7 (Table 3) and ASTER data (Table 5) have comparable accuracies in the estimation of most of the variables. Another similarity of
ASTER-based modelling with Landsat-based was the relative weakness
of R2 in indicating model performance. Nonetheless, differences between ASTER and Landsat derived estimations were observed in certain
instances. For example, ASTER performed better than Landsat in modelling pH, EC, Fe and U using NIR reflectance, while Landsat was better at
modelling EC, SO₄2−, Cl−, Na and K using SWIR 1 reflectance. One of the
advantages of ASTER over Landsat is the greater number of bands it possesses in the SWIR region; this allowed for the identification of bands
that provided the best estimation capability for certain variables in all
variable groups, such as Eh from field parameters and Ni from trace
elements (Table 4). Another advantage of having more SWIR bands
was the opportunity for better accuracy in some of these bands. For example, SWIR 6 was the best predictor for pH, Eh, Na and Cu while other
SWIR bands returned accuracies fairly comparable to the best
predictions.
Fig. 3 illustrates the correlations between estimated and observed
values of the same field and chemical properties that were presented
for Landsat-based estimations in Fig. 2. Eh was best modelled using reflectance of SWIR 6 band, compared to the blue band using Landsat data.
The green band was the best source of data for estimating NO−
3 . It
should be noted that ASTER does not have the blue band, which was
Please cite this article as: Tesfamichael, S.G., Ndlovu, A., Utility of ASTER and Landsat for quantifying hydrochemical concentrations in abandoned
gold mining, Sci Total Environ (2017), https://doi.org/10.1016/j.scitotenv.2017.09.335
8
S.G. Tesfamichael, A. Ndlovu / Science of the Total Environment xxx (2017) xxx–xxx
Table 4
A summary of AICc regression analysis results to estimate chemical variables using zonal statistics derived from individual bands of Landsat-5 acquired in November 2003. ΔAICc refers to
the differences between the model selected in this study and the best model with the lowest AICc value per band. RMSE (%) is the relative RMSE quantified as the ratio of absolute RMSE and
the observed mean chemical value multiplied by 100. Highlighted results show good accuracies based on RMSE followed by R2.
the best source of data for Landsat-based NO−
3 estimation. There was
slight decrease in the estimation accuracy of NO−
3 in that the deviations
of estimations were marginally higher, compared to those observed
using Landsat-based estimations (Fig. 2). In contrast, Ca was estimated
at better accuracy levels (R2 = 0.36; RMSE = 52%) than was using
Landsat (R2 = 0.27; RMSE = 56%). Cobalt, on the other hand, had
poorer estimation accuracy using ASTER than Landsat, although the
overall accuracy in both cases remained weak.
A comparison was made to assess whether or not the similarities between best case ASTER and Landsat-7 estimations of the four selected
variables (Figs. 2 and 3) were random. This was achieved by plotting estimation errors of the two images per variable; for example, estimation
errors of Eh were compared between best case Landsat blue band and
best case ASTER. Fig. 4 presents the results of the comparisons. This approach is robust since it determines the direction and magnitude of estimation errors per sample, compared to RMSE which aggregates errors
of multiple samples and thereby potentially concealing differences
among individual errors. A correlation of estimation errors that is
aligned perfectly along a line that passes through the origin as well as
bisects the 1st (+, +) and 3rd (−, −) quadrants indicates the ideal correspondence of errors and thus equivalence of estimation models. Accordingly, there were strong relative similarities in estimations (or
estimation errors) of Eh and Ca by ASTER and Landsat with R2 exceeding
0.60. In both cases, the fit lines almost bisect the 1st and 3rd quadrants
while errors were distributed along the lines well. Furthermore, few errors fell in the 2nd (−, +) and 4th (+, −) quadrants. In contrast, NO−
3
estimation between ASTER and Landsat varied considerably with
R2 ~ 0.44. This rather weak similarity is despite the fact that each had
the best estimation accuracy, as shown in Figs. 2 and 3. Nevertheless,
the fit line matches almost perfectly the ideal scenario, indicating similarity in most of the estimation errors while few isolated samples
displayed high error levels that resulted in low R2. The estimation errors
of Co using ASTER and Landsat had the worst correlations at an R2 of
0.26, and a considerable deviation from the ideal fit line scenario.
This is also confirmed by a large number of errors located in the 2nd
(−, +) and 4th (+, −) quadrants, even though most of those in the
1st (+, +) and 3rd (−, −) quadrants deviate significantly from the fit
line. Furthermore, the actual fit line deviates considerably from the
ideal fit line scenario. This weak correlation conforms to the poor estimation capability observed in Figs. 2 and 3.
4. Discussion
4.1. Prediction potential of remotely-sensed data
The impact of mining on environmental variables such as water is
immense with the effects felt in surface as well as ground ecosystems.
Quantifying and monitoring such problems is vital to implementing intervention strategies, but the traditional method of achieving this is expensive, time-consuming and requires following strict safety protocols.
Remote sensing based approaches have shown a great promise in the
assessment of water qualities associated with mining activities; however much of the work has been using qualitative classification approaches
(e.g. Torbick et al., 2013; Yucel et al., 2014; Riaza et al., 2015;
Rodríguez-Hernández et al., 2016; Snapir et al., 2017). Moreover, the
focus has mainly been on a limited number of water quality indicators.
It is against this background that this study sought to explore the potential of quantifying several geochemical properties of water bodies
Please cite this article as: Tesfamichael, S.G., Ndlovu, A., Utility of ASTER and Landsat for quantifying hydrochemical concentrations in abandoned
gold mining, Sci Total Environ (2017), https://doi.org/10.1016/j.scitotenv.2017.09.335
S.G. Tesfamichael, A. Ndlovu / Science of the Total Environment xxx (2017) xxx–xxx
9
Table 5
A summary of AICc regression analysis results to estimate chemical variables using zonal statistics derived from individual bands of ASTER acquired in January 2002. ΔAICc refers to the
differences between the model selected in this study and the best model with the lowest AICc value per band. RMSE (%) is the relative RMSE quantified as the ratio of absolute RMSE
and the observed mean chemical value multiplied by 100. Highlighted results show good accuracies based on RMSE followed by R2.
related to abandoned mining activities using moderate-resolution
remotely-sensed data as predictors. In doing so, it compared two
remotely-sensed data sources (Landsat and ASTER).
We opted to apply an information-theoretic approach, namely the
AIC, to build quantitative models that utilize statistical values (mean,
minimum, maximum, range, standard deviation and sum) of Landsat
and ASTER individual spectral bands within each delineated water
body (zone) as predicting variables. The AIC modelling was deemed
suitable in this study because of the high level of uncertainty in field observations of data such as those used in this study. Informationtheoretic approach allows for development of all potential models
from a given set of dependent and independent variables, without imposing statistical criteria such as satisfying a predetermined significance
level on the latter. We strongly believe that this approach is appropriate
for such observational data since it gives the opportunity to choose a
model that compromises accuracy and uncertainty due to sampling
error. This approach was applied to a limited extent by Bonansea et al.
(2015) for assessing chlorophyll and Secchi depth concentrations albeit
in a non-mining related environment. The results obtained through this
approach and using Landsat- and ASTER data in this study had mixed
success levels. Most field parameters as well as major anions and cations
were estimated fairly well in terms of RMSE, with comparable performance across spectral bands. Trace elements on the other hand were estimated quite poorly, despite the high R2 values for most of them
(Tables 3–5). Such a misleading R2 value unsupported by RMSE is the
product of a sample with an extreme value in these trace elements.
However the same sample did have values within the range of the population in terms of other chemical variables. It was therefore decided
Please cite this article as: Tesfamichael, S.G., Ndlovu, A., Utility of ASTER and Landsat for quantifying hydrochemical concentrations in abandoned
gold mining, Sci Total Environ (2017), https://doi.org/10.1016/j.scitotenv.2017.09.335
10
S.G. Tesfamichael, A. Ndlovu / Science of the Total Environment xxx (2017) xxx–xxx
Fig. 3. Comparison of observed and estimated values using ASTER bands. Redox potential (Eh) was estimated using spectral reflectance of SWIR 6; Nitrate (NO−
3 ) and Cobalt (Co) were
estimated using spectral reflectance of the green band; Calcium (Ca) was estimated using spectral reflectance of NIR. The dotted line in each figure shows the ideal one-to-one relationship.
not to exclude the sample from the analysis. These observations are not
surprising in view of the data distribution presented in Table 1, in which
the variables with low coefficient of variation have considerably different mean and median values indicating the level of skewness influenced
by extreme values.
Differences in sample values between 2002 and 2003 observed data
(Table 1) also influenced estimation accuracies obtained using Landsat
bands of the two dates as predictors (Tables 3 and 4). The discrepancy
between the best models per chemical category of the two dates was
expected for two reasons. Firstly, model performances were reflective
of data distributions with those closer to normal distributions (mean–
median equivalence as well as low CV) have a better accuracy. For example, NO−
3 had a better normal distribution in the 2002 dataset than
2003 the dataset, resulting in better estimation accuracy in the former
case; the opposite is true for K. Secondly, an AIC model stands given
both predictor and dependent variables from which it is developed
(Burnham and Anderson, 2002). Model variation is therefore expected
for different input data.
Previous studies targeted fewer water quality parameters, and is
therefore difficult to make direct comparisons with our findings. Furthermore, studies to develop prediction models that utilize remotelysensed spectra are rare. The study by Schroeter and Gläβer (2011), for
example, characterized selected water quality parameters including
pH and Fe (the only similarities with our study) using various statistical
techniques among which a bivariate correlation analysis was the only
quantitative method used to correlate the parameters with spectra.
Although significant, the correlations between green and red bands, respectively, with Fe were b 50%, while they were poorer and insignificant
with pH. Kopačková (2014) obtained up to an R2 of 0.76 while Ong and
Cudahy (2014) reported an R2 of 0.72 for estimating pH using airborne
hyperspectral data at abandoned lignite and pyrite mining, respectively.
These results are better than ours, however both are carried out on soil
compared to water bodies in our study. Besides, our study utilized low
spectral (b 10 bands in each of the ASTER and Landsat images) and spatial (15 m or 30 m) resolutions compared to high spectral (N 100 narrow
bands) and spatial (3 m or 5 m) resolutions used by these studies.
4.2. Comparison between ASTER and Landsat
The performances of ASTER and Landsat derived data in estimating
the variables of interest are generally comparable. This is encouraging
considering an almost three-month difference in the acquisition dates
of the images. It should however be noted that both months fall within
the same season in the study area, thus the effect of seasonal change on
the behavior of variables has been minimized. Moreover, the delineation process to specify the sampling zone around each sample point ensured that mostly pixels representing water bodies were specified. As a
result, pixels reflected from vegetation that varies with time were
avoided from the zones. However, it is still worth noting the likelihood
effect of the time difference between the two images. Although all the
data were acquired in the same season (summer), the Landsat data
(March 2002) was acquired closer to the field survey (April 2002)
than was the ASTER data (January 2002). Temporal coincidence between field (reference) data and remotely-sensed data plays a vital
role in terrestrial and aquatic feature characterization. Tutu et al.
(2008) reported the seasonal variation of the chemical properties in
the April 2002 versus the August 2003 data, and attributed the variation
to dissolution of efflorescent crusts and subsequent erosion into the
Please cite this article as: Tesfamichael, S.G., Ndlovu, A., Utility of ASTER and Landsat for quantifying hydrochemical concentrations in abandoned
gold mining, Sci Total Environ (2017), https://doi.org/10.1016/j.scitotenv.2017.09.335
S.G. Tesfamichael, A. Ndlovu / Science of the Total Environment xxx (2017) xxx–xxx
11
Fig. 4. Correspondence of error between Landsat and ASTER estimates. The comparison focused on the models and variables presented in Figs. 2 and 3.
water bodies as well as increase in the water table that leads to seepage
into streams and water bodies during or immediately after wet (summer) seasons. It is therefore logical to see a better comparison between
data with a shorter time gap, such as a 1-month difference between the
field and the Landsat data as opposed to a 4-month gap between the
field and the Landsat data. A dry season with very low cloud cover
and low vegetation vigor would be ideal for such a study, provided
that matching remotely-sensed and field data are available.
It is important to note the potential of having multiple narrow-range
SWIR bands, as observed in ASTER. Subdividing the broad SWIR 2 band
found in Landsat into narrower bands in ASTER resulted in the improved estimation accuracy in some of the bands. For example, Eh
was estimated best using SWIR 6 band, while a comparable accuracy
was achieved using SWIR 4. The SWIR is, for instance, known to be
suitable for defining water body extents (Schroeter and Gläβer, 2011),
and thus increasing the spectral qualities (through an increase in
spectral resolution and number of spectral bands) is logically expected
to improve the performance in the region of the electromagnetic
radiation. In comparison to ASTER and Landsat, recent satellite
missions such as WorldView and Sentinel-2 have fairly good spectral
characteristics that make them suitable candidates for water quality
monitoring.
The success of this study should also be viewed in lieu of the sampling protocol adopted in the study. In certain instances multiple samples were collected and treated independently within each water
body; this sampling design was followed to extract reflectance from
remotely sensed data. This approach is useful to capture the potential
variability that exists in a water body, as compared to treating an entire
water body as a homogenous sample by aggregating (averaging) multiple samples. Besides, the direct comparison between a point-based field
sample and a pixel that mixes information within a given area supported by the pixel size is inherently associated with errors. Despite this
problem, the accuracies obtained in this study using a pixel size of
15 m or larger are very encouraging. It is logical that a better spatial resolution would improve on the accuracies obtained in this study.
Nonetheless, underestimation of variables was observed in both
Landsat and ASTER images consistently. This is attributed to the inherent characteristics of the remotely-sensed data used in the study.
Broad-band multispectral data are known to saturate after certain concentration levels of earth features. The use of better spatial and spectral
resolution remotely-sensed data would improve estimation accuracy as
demonstrated by Ong and Cudahy (2014). In this regard, the continuous
technological improvement in such characteristics bodes well for affordable and accurate modelling of water contamination in mining
areas. For instance, WorldView imagery equipped with high spatial resolution (0.5 m) and up to 13 spectral bands has a great potential to mapping and monitoring water qualities. One of the advantages of this
imagery is the availability of a number of bands that are sensitive to
water characteristics. Similarly, Sentinel-2 multispectral imagery is
well equipped in terms of spectral and to a limited extent spatial resolution that can be improved by fusing with higher spatial resolution imagery (e.g. SPOT).
Please cite this article as: Tesfamichael, S.G., Ndlovu, A., Utility of ASTER and Landsat for quantifying hydrochemical concentrations in abandoned
gold mining, Sci Total Environ (2017), https://doi.org/10.1016/j.scitotenv.2017.09.335
12
S.G. Tesfamichael, A. Ndlovu / Science of the Total Environment xxx (2017) xxx–xxx
5. Conclusion
This study investigated the potential of moderate-resolution multispectral remotely-sensed data in quantifying several hydrochemical parameters in water bodies in the vicinity of abandoned gold mines.
Individual bands of ASTER and Landsat images were used as inputs to
estimate field and laboratory observed parameters in an AIC modelling
approach, which built all potential models from which the user with experience in the expected data structure of chemical properties can select
the appropriate ones. The results obtained in the study are mixed in
terms of accuracy. Field parameters including pH and Eh were estimated
at relatively good accuracy levels, along with most major anions and
cations, with RMSE values less than observed means (%RMSE b 100).
In contrast, lower accuracies were obtained for trace elements due to
high variability in sample values relative to sample size.
ASTER and Landsat images yielded comparable results in most cases,
despite the difference in acquisition time. The relative superiority of
Landsat (for example for NO−
3 ) is attributed to the temporal closeness
to the reference (field and laboratory) data. On the other hand, the comparability is also attributed to similarities in terms of spatial and spectral
resolutions of the two images, although ASTER's greater number of
SWIR bands should be explored in further studies. These performances
are encouraging considering that multispectral remote sensing systems
are freely available to the public, as opposed to expensive hyperspectral
remote sensing systems. The continuous improvement in spectral and
spatial resolutions of affordable data (e.g. WorldView, SPOT, Sentinel2 multispectral and imagery from unmanned aerial systems (UAS))
are expected to facilitate the design and the implementation of reliable
and fast monitoring systems that will assist mitigation efforts.
Acknowledgements
This research was funded by the University of Johannesburg. The National Research Foundation (NRF) of South Africa provided one of the
authors with a bursary.
References
Akaike, H., 1973. Information theory as an extension of the maximum likelihood principle. In: Petrov, B.N., Csaki, F. (Eds.), Second International Symposium on Information
Theory. Akademiai Kiado, Budapest, pp. 267–281.
Bonansea, M., Rodriguez, M.C., Pinott, L., Frerreo, S., 2015. Using multi-temporal Landsat
imagery and linear mixed models for assessing water quality parameters in Rio
Tercero reservoir (Argentina). Remote Sens. Environ. 15, 28–41.
Burnham, K.P., Anderson, D.R., 2002. Model Selection and Multimodel Inference: A Practical -Information-Theoretic Approach. Second edition. Springer-Verlag, New York.
Chander, G., Markham, B.L., Helder, D.L., 2009. Summary of current radiometric calibration for Landsat MSS, TM, ETM+ and EO-1 ALI sensors. Remote Sens. Environ. 113,
893–903.
Choudhury, B.U., Malang, A., Webster, R., Mohapatra, K.P., Verma, B.C., Kumar, M., Das, A.,
Islam, M., Hazarika, S., 2017. Acid drainage from coal mining: effect on paddy soil and
productivity of rice. Sci. Total Environ. 583, 344–351.
Davidson, C., 2003. Catchment Diagnostic Framework for the Klip River Catchment, Vaal
Barrage, October 1998–September 1999. Unpublished MSc Research Report. University of Witwatersrand, Johannesburg.
Davies, G.E., Calvin, W.M., 2016. Quantifying iron concentration in local and synthetic acid
mine drainage: a new technique using handheld field spectrometers. Mine Water Environ. https://doi.org/10.1007/s10230-016-0399-z.
Dube, T., Mutanga, O., Seutloali, K., Adelabu, S., Shoko, C., 2015. Water quality monitoring
in sub-Sharan African lakes: a review of remote sensing applications. Afr. J. Aquat. Sci.
40, 1–7.
Durand, J.F., 2012. The impact of gold mining on the Witwatersrand on the rivers and
karst system of Gauteng and North West province, South Africa. J. Afr. Earth Sci. 68,
24–43.
FEPA, 2011. Freshwater Ecosystem Priority Areas. South African National Biodiversity
Institute.
Handley, J., 2004. Historic Overview of the Witwatersrand Goldfields. First edition.
Handley, Howick, South Africa, pp. 73–116.
Hurvich, C.M., Tsai, C.-L., 1989. Regression and time series model selection in small samples. Biometrika 76, 297–307.
Johnson, D.B., Hallberg, K.B., 2005. Acid mine drainage remediation options: a review. Sci.
Total Environ. 338, 3–14.
Khalil, A., Hanich, L., Hakkou, R., Lepage, M., 2014. GIS-based environmental database for
assessing the mine pollution: a case study of an abandoned mine site in morocco.
J. Geochem. Explor. 144 (Part C468–477).
Kopačková, V., 2014. Using multiple spectral feature analysis for quantitative pH mapping
in a mining environment. Int. J. Appl. Earth Obs. Geoinf. 28, 28–42.
Kullback, S., Leibler, R.A., 1951. On information and sufficiency. Ann. Math. Stat. 22, 79–86.
Lottermoser, B.G., 2003. Mine Wastes. Chapter 3: Mine Water. Springer, New York City,
pp. 83–141.
McCarthy, T.S., 2011. The impact of acid mine drainage in South Africa. S. Afr. J. Sci. 107,
1–7.
McCarthy, T.S., Arnold, V., Venter, J., Ellery, W.N., 2007. The collapse of Johannesburg's Klip
River wetland. S. Afr. J. Sci. 103, 391–397.
McIntyre, N., Bulovic, N., Cane, I., McKenna, P., 2016. A multi-disciplinary approach to understanding the impacts of mines on traditional uses of water in Northern Mongolia.
Sci. Total Environ. 557–558, 404–414.
Naicker, K., Cukrowska, E., McCarthy, T.S., 2003. Acid mine drainage arising from gold
mining activity in Johannesburg, South Africa and environs. Environ. Pollut. 122,
29–40.
Ong, C.C.H., Cudahy, T.J., 2014. Mapping contaminated soils: using remotely-sensed
hyperspectral data to predict pH. Eur. J. Soil Sci. 65, 897–906.
Riaza, A., Buzzi, J., Garica-Melendez, E., Carrer, V.M., Sarmiento, A., Muller, A., 2015. Monitoring acidic water in a polluted river with hyperspectral remote sensing (HyMap).
Hydrol. Sci. J. 60, 1064–1077.
Rodríguez-Hernández, A., Briones-Gallardo, R., Razo, I., Noyola-Medrano, C., Lázaro, I.,
2016. Processing methodology based on ASTER data for mapping mine waste
dumps in a semiarid polysulphide mine district. Can. J. Remote. Sens. 42, 643–655.
Sakamoto, Y., Ishiguro, M., Kitagawa, G., 1986. Akaike information criterion statistics. KTK
Scientific Publishers, Tokyo, Japan.
Sams III, J.I., Veloski, G.A., 2003. Evaluation of airborne thermal infrared imagery for locating mine drainage sites in the Lower Kettle Creek and Cooks Run Basins, Pennsylvania, USA. Mine Water Environ. 22, 85–93.
Schroeter, L., Gläβer, C., 2011. Analyses and monitoring of lignite mining lakes in Eastern
Germany with spectral signatures of Landsat TM satellite data. Int. J. Coal Geol. 86,
27–39.
Scott, R., 1995. Flooding of Central and East Rand Gold Mines – An Investigation into Controls Over the Inflow Rate, Water Quality and the Predicted Impacts of Flooded Mines.
Report No. 486/1/95. Water Research Commission, Pretoria.
Shang, J., Morris, B., Howarth, P., Lévesque, J., Staenz, K., Neville, B., 2009. Mapping mine
tailing surface mineralogy using hyperspectral remote sensing. Can. J. Remote. Sens.
35, S126–S141.
Snapir, B., Simms, D.M., Waine, T.W., 2017. Mapping the expansion of galamsey gold
mines in the cocoa growingarea of Ghana using optical remote sensing. Int. J. Appl.
Earth Obs. Geoinf. 58, 225–233.
Torbick, N., Hession, S., Hagen, S., Wiangwang, N., Becker, B., Qi, J., 2013. Mapping inland
lake water quality across the Lower Peninsula of Michigan using Landsat TM imagery.
Int. J. Remote Sens. 34, 7607–7624.
Tutu, A., McCarthy, T.S., Cukrowska, E., 2008. The chemical characteristics of acid mine
drainage with particular reference to sources, distribution and remediation: the Witwatersrand Basin, South Africa as a case study. Appl. Geochem. 23, 3666–3684.
Urbanski, J.A., Wochna, A., Bubak, I., Grzybowski, W., Lukawska-Matuszewska, K., Łacka,
M., Sliwińska, S., Wojtasiewicz, B., Zajạczkowski, M., 2016. Application of Landsat 8
imagery to regional-scale assessment of lake water quality. Int. J. Appl. Earth Obs.
Geoinf. 51, 28–36.
Wang, F., Han, L., Kung, H.-T., van Arsdale, R.B., 2006. Applications of Landsat-5 TM imagery in assessing and mapping water quality in Reelfoot Lake, Tennessee. Int. J. Remote
Sens. 27, 5269–5283.
Yenilmez, F., Kuter, N., Emil, M.K., Aksoy, A., 2011. Evaluation of pollution levels at an
abandoned coal mine site in Turkey with the aid of GIS. Int. J. Coal Geol. 86 (2011),
12–19.
Yucel, D.M., Yucel, M.A., Baba, A., 2014. Change detection and visualization of acid mine
lakes using time series satellite image data in the geographical information system
(GIS): Can (Conakkale) Country, NW Turkey. Environ. Earth Sci. 72, 4311–4323.
Zhao, D., Cai, T., Jiang, H., Xu, D., Zhang, W., 2011. Estimation of water clarity in Taihu Lake
and surrounding rivers using Landsat imagery. Adv. Water Resour. 34, 165–173.
Please cite this article as: Tesfamichael, S.G., Ndlovu, A., Utility of ASTER and Landsat for quantifying hydrochemical concentrations in abandoned
gold mining, Sci Total Environ (2017), https://doi.org/10.1016/j.scitotenv.2017.09.335
Документ
Категория
Без категории
Просмотров
0
Размер файла
3 646 Кб
Теги
scitotenv, 2017, 335
1/--страниц
Пожаловаться на содержимое документа