Professional Documents
Culture Documents
Xiaoji Zeng
Zhifeng Liu
Chunyang He
Qun Ma
Jianguo Wu
Xiaoji Zeng, Zhifeng Liu, Chunyang He, Qun Ma, Jianguo Wu, “Detecting surface coal mining areas
from remote sensing imagery: an approach based on object-oriented decision trees,” J. Appl.
Remote Sens. 11(1), 015025 (2017), doi: 10.1117/1.JRS.11.015025.
Detecting surface coal mining areas from
remote sensing imagery: an approach based on
object-oriented decision trees
Abstract. Detecting surface coal mining areas (SCMAs) using remote sensing data in a timely
and an accurate manner is necessary for coal industry management and environmental assess-
ment. We developed an approach to effectively extract SCMAs from remote sensing imagery
based on object-oriented decision trees (OODT). This OODT approach involves three main
steps: object-oriented segmentation, calculation of spectral characteristics, and extraction of
SCMAs. The advantage of this approach lies in its effective integration of the spectral and spatial
characteristics of SCMAs so as to distinguish the mining areas (i.e., the extracting areas, stripped
areas, and dumping areas) from other areas that exhibit similar spectral features (e.g., bare soils
and built-up areas). We implemented this method to extract SCMAs in the eastern part of Ordos
City in Inner Mongolia, China. Our results had an overall accuracy of 97.07% and a kappa
coefficient of 0.80. As compared with three other spectral information-based methods, our
OODT approach is more accurate in quantifying the amount and spatial pattern of SCMAs
in dryland regions. © 2017 Society of Photo-Optical Instrumentation Engineers (SPIE) [DOI: 10
.1117/1.JRS.11.015025]
Keywords: surface coal mining area; object-oriented decision trees; remote sensing; Inner
Mongolia; Ordos.
Paper 16952 received Dec. 14, 2016; accepted for publication Mar. 13, 2017; published online
Mar. 23, 2017.
1 Introduction
Surface coal mining areas (SCMAs) refer to places where coal reserves are exploited using
opencast mining techniques, including the surrounding areas associated with mining activities.1
Specifically, SCMAs are composed of extracting areas (EAs), stripped areas (SAs), and dumping
areas (DAs).1 With increasing global coal production, the world has experienced dramatic
growth in SCMAs during the past century.1 Currently, SCMAs account for ∼40% of global
coal production.1
However, SCMAs worldwide are facing serious environmental challenges.1–4 For example,
SCMAs eradicate and damage native vegetation, resulting in loss of natural habitat and declines
in biodiversity.5,6 In some areas, dust and sulfide in the air, as well as heavy metals and toxic
substances in the water and soil, have significantly exceeded guidelines, leading to serious air,
water, and soil pollution in regions with SCMAs.2,3 In addition, opencast mining activities
destroy soil and geological structures, and have frequently caused serious soil erosion
and geological disasters, such as subsidence, landslides, and debris flows in these areas.1,3
Thus, timely and accurate detection of SCMAs are crucial for coal industry management and
environmental assessment.1,3,4
Since Chase and Pettyjohn7 used Landsat data to map SCMAs in Coshocton, Ohio, United
States, in 1972, remote sensing techniques have been widely used to detect SCMAs because of
their synoptic views and repeated imaging of large geographic areas.8–13 For example, Russell
et al.8 extracted SCMAs in southwestern Indiana, United States, using Landsat data. Mamula10
used aerial photographs and Landsat imagery to monitor SCMAs in northern Great Plains,
United States. Parks and Petersen14 quantified the spatial patterns of SCMAs in 1983 in central
Pennsylvania, United States. Fernandez-Manso et al.15 used remote sensing images to detect
SCMAs in United States, Australia, and Spain. Qian et al.16 extracted information about
the dynamics of SCMAs from 1978 to 2011 in Holingol City, China. LaJeunesse Connette17
analyzed the dynamics of SCMAs from 2002 to 2015 in Myanmar using Landsat data and
freely-available fine-resolution images of Google Earth.
At present, four remote sensing-based approaches are commonly used to detect SCMAs.
These include visual interpretation, supervised classification, decision trees, and object-oriented
classification. For instance, using visual interpretation, Prakash and Gupta5 extracted SCMAs
from 1975 to 1994 in the Jharia coalfield area of India. Based on supervised classification,
Demirel et al.18,19 obtained information on SCMAs in the Göynük region, Turkey, from
2004 to 2008. Decision trees were used in Zhai et al.,20 who identified SCMAs from 1979
to 2010 in Wuhai City, Inner Mongolia, China. Using object-oriented classification, Hou
and Hu21 extracted SCMAs in Jiaozuo city in Henan Province, China. However, visual
interpretation requires large amounts of human time and resources, thus not appropriate for
timely detection of SCMAs over broad spatial scales. The supervised classification methods,
decision trees, and object-oriented classification are all spectral information-based approaches,
which often result in relatively large errors when extracting SCMAs because SAs, DAs, barren
land, and built-up areas have similar spectral characteristics.18,19,22 To improve the accuracy of
detecting SCMAs with remote sensing data, new approaches are needed.
Object-oriented decision trees (OODT), which combine object-oriented techniques and
decision trees, provide a new way to extract SCMAs23 in a timely and more accurate manner.
The OODT uses segmented objects as the basic classification units, taking advantage of a variety
of information on spectra, shape, texture, and context that can improve the classification
accuracy.24 The OODT has already been successfully used to detect built-up areas, grasslands,
shrublands, forests, and water.23,24 For example, Laliberte et al.23 extracted the rangelands in the
Jornada del Muerto Basin. Qi et al.25 extracted the built-up area, barren land, vegetation, and
water in Guangzhou, China. However, the OODT has not been used to extract SCMAs.23–25
In this paper, we introduce a new OODT approach for extracting SCMAs. We first describe
the basic idea and the procedures involved in the new OODT approach. Then, we demonstrate
the new OODT approach through the detection of SCMAs in the eastern part of Ordos city,
a dryland region in Inner Mongolia, China, using Landsat-8 operational land imager (OLI) data.
Next, we assess the accuracy of our new approach against visually interpreted results of SCMAs
from high-resolution imagery and against three other remote sensing classification methods.
Finally, we discuss the advantages and limitations of the new OODT approach.
Fig. 1 The study area, located in Ordos City, Inner Mongolia, China.
parts: the EA, the SA, and the DA (Fig. 2). In this study, we lumped SA and DA together
(denoted as SA/DA) because these two parts were spectrally indistinguishable in many cases.
A Landsat-8 OLI multispectral image obtained from the United States Geological Survey
(USGS)’s website (Ref. 27) was used to extract the SCMAs. The Landsat-8 OLI image was
acquired on July 30, 2014, with a path and row of 127 and 33, respectively, and a spatial res-
olution of 30 m. This image contained spectral information from eight bands ranging from the
blue visible channel to the infrared channel.28 In addition, a Gaofen-1 image obtained from
the Chinese Academy of Sciences’ Remote Sensing Cloud Platform (Ref. 29) was used to assess
Fig. 2 Illustration of the spatial structure of SCMAs, consisting of the EAs, SAs, and DAs. (a) An
example of a SCMA in the Gaofen-1 image (R: Band 1, G: Band 2, and B: Band 3). (b) An example
of a SCMA in the Landsat-8 OLI image (R: Band 6, G: Band 5, and B: Band 4).
Fig. 3 Flow chart of our method, showing the major steps of implementation of our OODT
approach.
the accuracy of our results. The Gaofen-1 image was acquired on June 12, 2014, with a spatial
resolution of 2 m. The Gaofen-1 image has one panchromatic band (0.45 to 0.90 μm) and four
multispectral bands (0.45 to 0.52 μm, 0.52 to 0.59 μm, 0.63 to 0.69 μm, and 0.77 to 0.89 μm).30
2.2 Methods
Our OODT approach involves three main steps: object-oriented segmentation, calculation of
spectral characteristics, and extraction of SCMAs (Fig. 3). This approach is not entirely new,
but rather, it is based on other existing methods, such as supervised classification, decision trees,
and object-oriented classification and tailored particularly for detecting surface mining areas in
dryland regions.
shape, smoothness, and compactness. In accordance with the methods suggested by Laliberte
et al.,23 we used the Landsat-8 OLI image as the input data and used visual interpretation to
determine the optimal segmentation parameters. These parameters were selected based on
the values at which most pixels in an EA can be merged into an object. Specifically, the
scale parameter was set to 200, and the weights of color, shape, smoothness, and compactness
were set to 0.9, 0.1, 0.5, and 0.5, respectively. Using these parameters, we segmented the
Landsat-8 OLI image and merged the pixels into objects.
where B2, B4, B5, and B6 denote band 2, band 4, band 5, and band 6 in the Landsat-8 OLI
image, respectively.
1
EAi ¼ ; (4)
0 Otherwise
where EAi is the class value for the i’th object, with the value of 1 indicating an EA and 0 is
non-EA object; NDCIi and BAIi are the NDCI and BAI class values for the i’th object; and
T NDCI and T BAI are the optimal thresholds of NDCI and BAI for identifying EAs, respectively.
Then, we identified potential SAs or DAs from the non-EA objects using the following
logical function:
NDCIi > T NDCI and BAIi < T BAI and NDCIi < T NDCI
EQ-TARGET;temp:intralink-;e005;116;188
1
PSDi ¼ ; (5)
0 Otherwise
where PSDi is the class value for the i’th object, with the value of 1 indicating a potential SA/DA
and 0 is a non-SCMA object; NDVIi is the value of the NDVI for the i’th object; and T NDVI is
the optimal threshold of NDVI for identifying a potential SA/DA.
Following Chen et al.34 and Liu et al.,35 we determined the optimal thresholds of the NDCI,
NDVI, and BAI by selecting those values that yielded results that best matched a selected set of
reference samples. Specifically, we visually selected a total of 700 samples: 100 samples for each
land use/cover type (i.e., EA, SA/DA, grassland, barren land, built-up area, river, and lake) from
the 17,797 segmented objects in the Landsat-8 OLI image. Through a trial and error experiment
using these samples, we determined the optimal thresholds for NDCI, NDVI, and BAI to be
−0.01, 0.13, and −0.05, respectively. In other words, these values of optimal thresholds,
when used in Eqs. (4) and (5), produced the most accurate classification of EAs, SAs/DAs,
and barren lands.
Theoretically, a greater number of samples should lead to a greater accuracy of the threshold
values, but too large number of samples would be too laborious, reducing the efficiency of
the method. The adequacy of sample size can be determined by the accuracy of classification
results using these threshold values. In our study, the sufficiently high accuracy of classification
results confirmed that our a priori choice of 700 samples was reasonable.
In the second step, we differentiated the real SAs/DAs from the potential SAs/DAs based on
the expected spatial adjacency relationship among EAs, SAs, and DAs as empirically observed
(Fig. 2). Specifically, all the objects surrounding EAs were considered potential SAs/DAs and
examined one by one to identify their neighboring objects. If a potential SA/DA adjoined one or
more EAs, then it was determined as a real SA/DA. These operations were performed iteratively
using the following logical equations:
8 P neighbor
<1
EQ-TARGET;temp:intralink-;e006;116;532
EAi;j >0
RSDi;t ¼ j ; ðt ¼ 0Þ ; (6)
:
0 Otherwise
8 P
< 1 RSDi;t−1 ¼ 1 RSDneighbor
EQ-TARGET;temp:intralink-;e007;116;469
i;j;t−1 > 0
RSDi;t ¼ j ; ðt ¼ 1;2; : : : ; nÞ; (7)
:
0 Otherwise
where RSDi;t is the class value for the i’th object at the t’th iteration (1 ¼ real SA∕DA and
0 ¼ non-SA∕DA); EAneighbori;j is the class value for the j’th object that adjoins the i’th object,
i.e., sharing a boundary or a node of the i’th object (1 ¼ EA and 0 ¼ non-EA); and RSDii;t−1 and
RSDneighbor
i;j;t−1 are the same terms at the (t − 1)’th iteration.
After all the SAs/DAs were classified, they were merged with the EAs identified earlier to
map out the SCMAs across the entire study area in 2014.
3 Results
Fig. 4 Schematic representation of accuracy assessment. (a) The original Gaofen-1 image of the
study area with a spatial resolution of 2 meters; (b) visually identified SCMAs from the Gaofen-1
image; (c) SCMAs classified by the proposed OODT method; and (d) areas of agreement and
disagreement between the results from the two methods.
Reference (km2 )
Note: Quantity disagreement is defined as “the amount of difference between the reference map and a com-
parison map that is due to the less than perfect match in the proportions of the categories.”36
Allocation disagreement is defined as “the amount of difference between the reference map and a comparison
map that is due to the less than optimal match in the spatial allocation of the categories, given the proportions of
the categories in the reference and comparison maps.”36
Reference (km2 )
Fig. 5 The distribution of SCMAs in the study area. (a) The spatial distribution of SCMAs; (b) the
size distribution of SCMAs; (c) the size distribution of EAs; and (d) the size distribution of combined
SAs/DAs.
Fig. 6 Comparison of the spatial extent and distribution of SCMAs extracted using different
methods.
10 km2 , accounting for 14.76% of the total area of SCMAs [Fig. 5(b)]. The largest SCMA had
an area of 16.57 km2 , accounting for 9.18% of the total area of SCMAs. Forty-six SCMAs had
an area between 1 and 10 km2 . These SCMAs had a total area of 111.09 km2 , accounting for
61.54% of the total area of SCMAs. In addition, nearly a half of the SCMAs (135) were smaller
than 0.1 km2 [Fig. 5(b)]. The smallest SCMA had an area of only 0.01 km2.
The total area of EAs was 55.06 km2 , accounting for 30.50% of the total SCMAs. Ten
SCMAs had EAs larger than 1 km2 , totaling 19.49 km2 and accounting for 35.40% of the
total area of EAs [Fig. 5(c)]. The largest EA was 3.81 km2 in area (6.92% of the total area
of EAs). In addition, 178 of the SCMAs (64.26%) had EAs smaller than 0.1 km2 , with the
smallest EA being 0.004 km2 [Fig. 5(c)].
The SAs and DAs together accounted for 125.47 km2 or 69.50% of the total area of SCMAs.
Only one SCMA had SAs/DAs larger than 10 km2 [Fig. 5(d)], with an area of 12.78 km2 (or
10.18% of the total area of combined SAs and Das). Twenty-nine SCMAs had SAs/DAs between
1 and 10 km2 . These SAs/DAs had a total area of 75.28 km2 , accounting for 60.00% of the
total area of combined SAs and DAs. In addition, 118 (42.60% of the SCMAs) had SAs/DAs
smaller than 0.1 km2 [Fig. 5(d)].
3. Build the decision tree based on red, green, and blue bands.
Fig. 7 Accuracy assessment of the largest and smallest SCMAs using different methods.
Acknowledgments
We would like to thank the anonymous reviewers for their valuable comments and suggestions.
This work was supported in part by the National Basic Research Program of China (Grant
Nos. 2014CB954303 and 2014CB954302) and the National Natural Science Foundation of
China (Grant Nos. 41621061 and 41501195).
References
1. World Coal Institute, “The coal resource—a comprehensive overview of coal,” 2009, World
Coal Institute, http://www.worldcoal.org (10 February 2016).
2. C. S. Rathore and R. Wright, “Monitoring environmental impacts of surface coal mining,”
Int. J. Remote Sens. 14, 1021–1042 (1993).
3. M. A. Palmer et al., “Science and regulation. Mountaintop mining consequences,”
Science 327, 148–149 (2010).
4. J. Wu et al., “Historical landscape dynamics of Inner Mongolia: patterns, drivers, and
impacts,” Landscape Ecol. 30, 1579–1598 (2015).
5. A. Prakash and R. P. Gupta, “Land-use mapping and change detection in a coal mining area
—a case study in the Jharia coalfield, India,” Int. J. Remote Sens. 19, 391–410 (1998).
6. A. Erener, “Remote sensing of vegetation health for reclaimed areas of Seyitömer open cast
coal mine,” Int. J. Coal Geol. 86, 20–26 (2011).
7. P. E. Chase and W. Pettyjohn, “ERTS-1 investigation of ecological effects of strip mining
in eastern Ohio,” in Proc. of Conf. on Significant Results Obtained from ERTS-1,
1(Sect. A & B), pp. 561–568 (1973).
8. O. R. Russell et al., “Applications of ERTS-1 and aircraft imagery to mined land investi-
gations: Tullahoma, Tennessee University Space Institute,” in Remote Sensing of Earth
Resource Symp., pp. 1095–1105 (1973).
9. J. L. Solomon, W. F. Miller, and D. A. Quattrochi, “Development of a tree classifier
for discrimination of surface mine activity from Landsat digital data,” in Proc. of the
Annual Meeting of the American Society of Photogrammetry, Vol. 2, pp. 607–613 (1979).
10. N. Mamula, “Remote-sensing methods for monitoring surface coal mining in northern great
plains,” J. Res. US Geol. Surv. 6, 149–160 (1978).
11. E. T. Slonecker and M. J. Benger, “Remote sensing and mountaintop mining,” Remote Sens.
Rev. 20, 293–322 (2001).
12. L. Schroeter and C. Glasser, “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
(2011).
13. N. M. Saadi et al., “Integrating data from remote sensing, geology and gravity for geological
investigation in the Tarhunah area, northwest Libya,” Int. J. Digital Earth 1, 347–366
(2008).
14. N. F. Parks and G. W. Petersen, “High-resolution remote-sensing of spatially and spectrally
complex coal surface mines of central Pennsylvania—a comparison between simulated
SPOT MSS and landsat-5 thematic mapper,” Photogramm. Eng. Remote Sens. 53, 415–420
(1987).
15. A. Fernandez-Manso, C. Quintano, and D. Roberts, “Evaluation of potential of multiple
endmember spectral mixture analysis (MESMA) for surface coal mining affected area
mapping in different world forest ecosystems,” Remote Sens. Environ. 127, 181–193 (2012).
16. T. N. Qian et al., “Spatial-temporal analyses of surface coal mining dominated land deg-
radation in Holingol, Inner Mongolia,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.
7, 1675–1687 (2014).
17. K. J. LaJeunesse Connette et al., “Assessment of mining extent and expansion in Myanmar
based on freely-available satellite imagery,” Remote Sens. 8, 912 (2016).
18. N. Demirel, S. Duzgun, and M. K. Emil, “Landuse change detection in a surface coal mine
area using multi-temporal high-resolution satellite images,” Int. J. Min. Reclam. Environ.
25, 342–349 (2011).
19. N. Demirel, M. K. Emil, and H. S. Duzgun, “Surface coal mine area monitoring using
multi-temporal high-resolution satellite imagery,” Int. J. Coal Geol. 86, 3–11 (2011).
20. M. Zhai et al., “Remote sensing monitoring of the ecological environment in Wuhai mining
area since 1979,” Remote Sens. Technol. Appl. 27(6), 933–940 (2012).
21. F. Hou and Z. Hu, “Remote sensing information extraction of typical surface objects in
a coal mining area based on multiple-scale segmentation,” Bull. Surv. Mapp. 27(1), 22–25
(2012).
22. Y. Mao et al., “Study and validation of a remote sensing model for coal extraction based on
reflectance spectrum features,” Can. J. Remote Sens. 40, 327–335 (2014).
23. A. S. Laliberte, E. L. Fredrickson, and A. Rango, “Combining decision trees with hierar-
chical object-oriented image analysis for mapping arid rangelands,” Photogramm. Eng.
Remote Sens. 73, 197–207 (2007).
24. T. Blaschke, “Object based image analysis for remote sensing,” ISPRS J. Photogramm.
Remote Sens. 65, 2–16 (2010).
25. Z. X. Qi et al., “Integration of polarimetric decomposition, object-oriented image analysis,
and decision tree algorithms for land-use and land-cover classification using RADARSAT-2
polarimetric SAR data,” Photogramm. Eng. Remote Sens. 78, 169–181 (2012).
26. D. Xu et al., “Assessing the relative role of climate change and human activities in sandy
desertification of Ordos region, china,” Sci. China Ser. D-Earth Sci. 52, 855–868 (2009).
27. United States Geological Survey (USGS)'s, Earth Resources Observation and Science
Center (EROS), “Global visualization viewer,” http://glovis.usgs.gov/ (10 February 2016).
28. D. P. Roy et al., “Landsat-8: science and product vision for terrestrial global change
research,” Remote Sens. Environ. 145, 154–172 (2014).
29. Chinese Academy of Sciences (CAS), “Remote sensing cloud platform,” http://www.
rscloudmart.com/ (10 February 2016).
30. P. Hao, L. Wang, and Z. Niu, “Potential of multitemporal Gaofen-1 panchromatic/ multi-
spectral images for crop classification: case study in Xinjiang Uygur autonomous region,
China,” J. Appl. Remote Sens. 9(1), 096035 (2015).
31. M. Baatz and A. Schape, “Multiresolution segmentation: an optimization approach for
high quality multi-scale image segmentation,” in Angewandte Geographische Informations-
Verarbeitung XII, J. Strobl, T. Blaschke, and G. Griesbner, Eds., Vol. 12, pp 12–23,
Wichmann Verlag, Karlsruhe, Germany (2000).
32. C. J. Tucker, “Red and photographic infrared linear combinations for monitoring vegeta-
tion,” Remote Sens. Environ. 8, 127–150 (1979).
33. M. Bouziani, K. Goïta, and D.-C. He, “Automatic change detection of buildings in urban
environment from very high spatial resolution images using existing geodatabase and prior
knowledge,” ISPRS J. Photogramm. Remote Sens. 65, 143–153 (2010).
34. J. Chen et al., “Land-use/land-cover change detection using improved change-vector
analysis,” Photogramm. Eng. Remote Sens. 69, 369–379 (2003).
35. Z. Liu et al., “Extracting the dynamics of urban expansion in China using DMSP-OLS
nighttime light data from 1992 to 2008,” Landscape Urban Plann. 106, 62–72 (2012).
36. R. G. Pontius and M. Millones, “Death to kappa: birth of quantity disagreement and allo-
cation disagreement for accuracy assessment,” Int. J. Remote Sens. 32, 4407–4429 (2011).
Xiaoji Zeng is a PhD candidate at Beijing Normal University, China. He received his MS degree
in geography from Xinjiang University in 2011. His research interests include detection of
surface coal mining areas and assessment of their environmental impacts.