Spatial Prediction of Ground Subsidence Susceptibility Using an ...

Environmental Management (2012) 49:347–358 DOI 10.1007/s00267-011-9766-5

Spatial Prediction of Ground Subsidence Susceptibility Using an Artificial Neural Network Saro Lee • Inhye Park • Jong-Kuk Choi

Received: 8 March 2011 / Accepted: 24 September 2011 / Published online: 18 October 2011 Ó Springer Science+Business Media, LLC 2011

Abstract Ground subsidence in abandoned underground coal mine areas can result in loss of life and property. We analyzed ground subsidence susceptibility (GSS) around abandoned coal mines in Jeong-am, Gangwon-do, South Korea, using artificial neural network (ANN) and geographic information system approaches. Spatial data of subsidence area, topography, and geology, as well as various ground-engineering data, were collected and used to create a raster database of relevant factors for a GSS map. Eight major factors causing ground subsidence were extracted from the existing ground subsidence area: slope, depth of coal mine, distance from pit, groundwater depth, rock-mass rating, distance from fault, geology, and land use. Areas of ground subsidence were randomly divided into a training set to analyze GSS using the ANN and a test set to validate the predicted GSS map. Weights of each factor’s relative importance were determined by the backpropagation training algorithms and applied to the input factor. The GSS was then calculated using the weights, and GSS maps were created. The process was repeated ten times to check the stability of analysis model using a

S. Lee (&) Geoscience Information Center, Korea Institute of Geoscience & Mineral Resources, (KIGAM), 92, Gwahang-no, Yuseong-gu, Daejeon 305-350, Korea e-mail: [email protected] I. Park Department of Geoinformatics, University of Seoul, Siripdae-gil 13, Dongdaemun-gu, Seoul 130-743, Republic of Korea J.-K. Choi Korea Ocean Satellite Centre, Korea Ocean Research & Development Institute, 454 Haean-no, Sangrok-gu, Ansan, Gyeonggi 426-744, Korea

different training data set. The map was validated using area-under-the-curve analysis with the ground subsidence areas that had not been used to train the model. The validation showed prediction accuracies between 94.84 and 95.98%, representing overall satisfactory agreement. Among the input factors, ‘‘distance from fault’’ had the highest average weight (i.e., 1.5477), indicating that this factor was most important. The generated maps can be used to estimate hazards to people, property, and existing infrastructure, such as the transportation network, and as part of land-use and infrastructure planning. Keywords GIS  Abandoned underground coal mine  Ground subsidence  Artificial neural networks  Korea

Introduction The coal industry played an important role in the development of the South Korean economy and industry in the 1960s and 1970s. However, the coal industry began to decline in the 1980s along with the decrease in international oil prices. In 1988, Jeongseon, Gangwon-do, the largest coal-mining region in South Korea, had 173 coal mines. Today, none of these mines is operational because all of them were abandoned after implementation of a coalindustry rationalization action by the Korean government. Nationwide, only seven of 345 coal mines were operating as of 2005 (Ministry of Knowledge Economy 2007). Mine closure included no measures to protect against environmental damage, and risks associated with abandoned mines have been increasing over time. Abandoned mines produce [140,000 tons of wastewater/day in South Korea. Moreover, approximately 13 million km2 of land is contaminated with trace minerals and mine waste. Various heavy

123

348

metals flow from mine waste heaps in leachate, causing serious pollution in rivers and soil. However, little is known of the extent of basement coal pit and ground subsidence because most abandoned coal mines are located in sparsely populated mountain areas. Underground subsidence can cause damage to surface structures, such as railroad tracks, roads, and buildings, as well as human injury (Mine Reclamation Corporation 2007). Complete restoration of subsidence areas is difficult and costly. Moreover, ground subsidence is generally treated by simple reinforcements after subsidence has occurred. Therefore, general and systematic predictions and management plans for ground subsidence occurrence are necessary (Oh and others 2011). This study assessed and predicted discontinuous residual subsidence to produce a ground subsidence susceptibility (GSS) map of an area near abandoned underground coal mines using an artificial neural network (ANN) in a geographic information system (GIS) environment. GIS approaches provide a way to introduce information and knowledge from other data sources into the decisionmaking process, and they aid in the handling and manipulation of classified remote-sensing data (Adinarayana and Krishna 1996). The use of a GIS enables quantitative assessment of the consequences of heterogeneity in environmental systems over a broad range of spatial and temporal scales. Systematic integration of several surface features that indicate ground subsidence hazard is an important aspect in land-management studies. A database designed to support ground-subsidence decisions must contain various types of thematic information because of the interdisciplinary nature of ground-subsidence problems. Recent studies have examined GSS based on geological and geotechnical investigations as well as probability, statistical, fuzzy algebra, and ANN models in tandem with GIS applications (Ambrozˇicˇ and Turk 2003; Choi and others 2010; Esaki and others 2008; Kim and others 2006, 2009; Lee and others 2010; Mancini and others 2009; Oh and Lee 2010, 2011; Oh and others 2011; Quanyuan and others 2009; Turer and others 2008). Some GSS assessments have identified areas at high risk of subsidence. For example, Ambrozˇicˇ and Turk (2003) and Kim and others (2009) applied ANN models to predict ground subsidence. Using a probabilistic model, Zahiri and others (2006) applied the weights-of-evidence technique to derive rockfall potential associated with mining-induced subsidence. Kim and others (2006) and Oh and Lee (2010) assessed spatial GSS using a GIS with models based on frequencyratio and weights-of-evidence models. Oh and others (2011) applied probabilistic-based sensitivity analysis to determinate the effect of input factors on GSS distribution, and Esaki and others (2008) used a stochastic model to predict subsidence in a coal mining area. Mancini and

123

Environmental Management (2012) 49:347–358

others (2009) applied a multicriteria decision model to analyze salt-mining activities. As a statistical model, Lee and others (2010) applied logistic regression, whereas Choi and others (2010) constructed subsidence-susceptibility maps based on fuzzy relations for an abandoned underground coal mine. Oh and Lee (2011) integrated the GSS maps obtained using various models. Most previous studies have constructed GSS maps using the above-mentioned models. The present study assessed and predicted GSS using raster databases, in an ArcGIS grid format, of topographic, geologic, and geotechnical data and the locations of subsidence areas already discovered in the study area. ArcGIS 9.3 software (ESRI, CA) was used for database construction, coordinate conversion, grid production, overlay analysis, and spatial analysis. Using the major factors, GSS maps were drawn by application of ANN models and then validated by area-under-the-curve (AUC; Lee and Dan 2005) analysis and a field survey. By this approach, the major influences on ground subsidence were determined in a limited 1-km2 region, and a method for predicting GSS efficiently was established. Figure 1 presents a flowchart of the method employed in this research.

Data Study Area In choosing a study area, field investigations and reports relating to ground subsidence were carefully considered. The study site, Jeong-am (37°120 000 –37°130 000 N, 128°530 1000 –128°540 1000 E; see Fig. 2), was an important coal mining area and still has many cavities remaining from mining. Thus, areas of likely ground subsidence exist in Jeong-am. Local road no. 38 passes along the study area and is currently being expanded. The study area lies between Mt. Baek-Wu to the southwest and Mt. Ham-Beak to the southeast and is located on a developed ridge connecting the two mountains. Almost all coal in South Korea is anthracite, 85% of which was deposited during the upper Paleozoic era and the lower Mesozoic era in the Jangseong Formation of the Pyeongan Supergroup (Geological Society of Korea 1999). The Pyeongan Supergroup is divided into two parts in the Permian unit: the Sadong Group, consisting of the Bamchi and Jangseong formations, and the Gobangsan Group, consisting of the Hambaeksan, Tosagok, and Kohan formations. The Gobangsan Group locally overlies the Sadong Group conformably. Surface geology around the study area consists of the Makkol Formation, which is part of the Great Limestone Group (Choseon Supergroup), and the Manhang, Geumcheon, Jangseong, Hambaegsan, Dosagog, Kohan, and Donggo

Environmental Management (2012) 49:347–358

349

Fig. 1 Flow chart

formations of the Pyeongan Supergroup. The Jangseong Formation includes several coal beds, one of which has workable quality and thickness (Geological Society of Korea 1999). Coal mining in this area occurred from 1967 until 1989, mainly in the Samtan Gallery of the Jangseong Formation. Coal seams in this area typically have steep slopes ([60–70°). Average seam thickness was *1.3–2.5 m, although rich seams reached 4–15 m in thickness (Coal Industry Promotion Board 2005). Disused drifts are found from the central part of the study area toward the northeastern part. Most of these drifts range from 70 to 260 m in depth and tend to deepen from the center to the northeastern part of the study area, along the direction of the Jangseong Formation dip. Drifts are

centered mainly where the Jangseong Formation is exposed to the surface or where the Hambaegsan Formation covers the Jangseong Formation. The Jangseong Formation is composed mainly of alternating sandstone and shale deposits, with the shale having intercalations of two to three coal bed seams. The Hambaegsan Formation, an upper stratum of the major coal drifts, is composed of coarsely grained meta-sandstone and gray shale and is relatively resistant to weathering. However, the development of cleavages from severe folding has led to weaknesses in the rock masses of this formation. Mining operations are inferred to have been conducted near the surface (Coal Industry Promotion Board 1997). Although subsidence has not been reported in adjacent residential

123

350

Fig. 2 Study area with ground-subsidence locations

123

Environmental Management (2012) 49:347–358

Environmental Management (2012) 49:347–358

351

Fig. 3 Mechanism of ground subsidence of an underground cavity

areas, severe ground subsidence has occurred in mountainous areas near public facilities. A local road (no. 38) shows features of typical sinkhole collapse and deformations and cracks in the road surface. Several soil surface collapses were reported on arable lands near a flood control reservoir in the central–northern part of the study area (Coal Industry Promotion Board 2005). The total area of ground subsidence within the study area is *3,296 m2. Subsidence locations are depicted on topographic maps and in satellite images of the study area (Fig. 2). Ground subsidence can occur in areas of past underground mining activity. In underground mines, ground subsidence develops from the mine roof to the ground surface. Mine collapse with time is attributable to decreased shear strength, groundwater injection, and increased seepage force after coal mining. Figure 3 illustrates the general progression of subsidence due to an underground cavity. Input Factors Many studies have identified important factors¯including the depth and height of mine cavities, excavation method, degree of inclination of the excavation, scope of mining, structural geology, flow of groundwater, and the mechanical characteristics represented by the rock-mass rating (RMR)—that contribute to ground subsidence around coal mines (e.g., Waltham 1989; Coal Industry Promotion Board 1997). In this study, locations of ground subsidence and factors governing the occurrence of ground subsidence were collected in a vector-type spatial database and then represented on a grid using the ArcGIS software package. The spatial database is listed in Table 1. The database included a 1:5,000 ground-subsidence map, a 1:50,000 geological map, a 1:5,000 topographic map, a 1:5,000 land-use map, a 1:1,200 mine-tunnel map, borehole data, and satellite imagery with 1-m resolution. Jeong-am was chosen as the

Table 1 Spatial database selected for this study Map classification

Subclassification

Data type

Scale

Ground-subsidence

Subsidence area

Polygon

1:5,000

Topographic

Road

Line

1:5,000

Railroad

Line

River

Polygon

Building

Polygon

District

Polygon

Contour

Line

Land-use

Land use

Polygon

Borehole

Borehole

Point

1:5,000

Geological

Geology Fault

Polygon Line

1:50,000

1:5,000

Mining tunnel

Mining tunnel

Polygon

1:1,200

Satellite image

Satellite image

Image

1-m Resolution

site for investigation. Reliable accuracy of the spatial database is indispensable in a GIS environment. For this reason, accurate maps authorized by national organizations, such as the Coal Industry Promotion Board (for ground subsidence), the National Geographic Information Institute (for topography and land use), the Mine Reclamation Corporation (for mine tunnels and boreholes), and the Korea Institute of Geoscience and Mineral Resources (for geology), were collected while the scales of these maps were different. As shown in Fig. 4, eight variables extracted from the constructed spatial database were considered as factors of ground subsidence when calculating probability. Contours (5-m intervals) and survey base points of elevation were extracted from the topographic map, and a triangulated irregular network was made using the elevations. The slope angle was obtained from the digital elevation map. Areas of buildings, mountains, railways, fields, rivers, complex area, roads, and multi-purpose area use were extracted from the land-use map. Most studies have identified the

123

352 Fig. 4 Input factors for GSS mapping

123

Environmental Management (2012) 49:347–358

Environmental Management (2012) 49:347–358

353

Fig. 4 continued

123

354

scope of mine cavities as the major factor in ground subsidence. Therefore, constructing a database of mine cavity depth and width was an important step and was achieved using 1:1,200 mine-tunnel map and DEM calculated from the 1:5,000 topographic map as follows. GPS measurements were used to determine the exact positions of mine heads. These positions were then used to vectorize a hard copy of the mine-tunnel map. The mine cavity depth was defined as the difference between mine-tunnel elevation and DEM. The width was obtained by digitizing of width attribute of mine-tunnel map. Finally, the vectorized mine-tunnel map was converted to an ASCII grid file and subtracted from the digital elevation-map raster data. There were 13 boreholes at the study site, but RMR values and groundwater levels were not available for four of these boreholes. RMR is a geomechanical classification system for rocks developed by Z. T. Bieniawski between 1972 and 1973 (Bieniawski 1989). RMR has been constructed by classification bedrock according to uniaxial compressive strength of rock material, rock-quality designation, spacing of discontinuities, condition of discontinuities, and groundwater conditions. Then it was corrected for the orientation of discontinuities. An inverse-distance weighting interpolation was used to contour groundwater level, and geological data were extracted from the 1:50,000 geological map. The fault was detected from IKONOS imagery with 1-m resolution by a structural geologist. The distance from the lineament was calculated at 1-m intervals using the Euclidean-distance method. The factors were then Fig. 5 Architecture of the ANN

123

Environmental Management (2012) 49:347–358

represented by 2 9 2 m cells in 179 rows and 361 columns; the entire study area comprised 63,677 cells, and ground subsidence had occurred in 824 cells.

Methodology Theory An ANN is a ‘‘computational mechanism able to acquire, represent, and compute a mapping from one multivariate space of information to another, given a set of data representing that mapping’’ (Garrett 1994, p. 129). The backpropagation training algorithm is the most frequently used neural network method and was used in this study. It is trained using a set of examples of associated input and output values. The purpose of an ANN is to build a model of the data-generated weighting process so that the network can generalize and predict outputs from inputs that it has not previously seen. This learning algorithm is a multilayered neural network that consists of an input layer, hidden layers, and an output layer. The hidden- and outputlayer nodes process their inputs by multiplying them by a corresponding weight, summing the products, and processing the sum using a nonlinear transfer function. An ANN ‘‘learns’’ by adjusting the weights between the nodes in response to errors between the actual and target output values. At the end of this training phase, the neural network provides a model that should be able to predict a target value from a given input value.

Environmental Management (2012) 49:347–358

355

GSS were selected as training sites. Areas where ground subsidence has not occurred were classified as ‘‘areas not prone to ground subsidence,’’ and 50% of areas where ground subsidence was known to have occurred were assigned to the ‘‘areas prone to ground subsidence’’ training set. The remaining 50% of ground-subsidence areas were classified into an ‘‘areas prone to ground subsidence’’ data set that was used for validation. The back-propagation algorithm was then applied to calculate the weights between the input and hidden layers and between the hidden and output layers. A three-layered, feedforward network was implemented using the MATLAB software package (Mathworks, MA). The number of hidden layers and the number of nodes in a hidden layer required for a particular classification problem are not easy to deduce. In this study, an 8 9 16 9 1 structure was selected for the network, and input data was normalized in the range of 0.1–0.9. The nominal and interval class group data were converted to continuous values ranging between 0.1 and 0.9. Therefore, the continuous values were nominal, not ordinal, data, and the numbers denote the classification of the input data. The learning rate was set at 0.01, and the initial weights were randomly set at values between 0.1 and 0.3. The weights calculated from ten test cases were compared with determine whether the variation in the final weights depended on the selection of the initial weights. The weights between layers acquired by neural network training were calculated in reverse, and the contribution or importance of each factor was calculated. Weights that represent the contribution or importance of each factor were determined. A program developed by Hines (1997) was used for weight calculation; and for interpretation of the weight, we used a newly developed program employing MATLAB software. The model was trained for 5,000 epochs, and the root mean-square error (RMSE) value used for the stopping criterion was set at 0.01. If this RMSE value was not

Two stages are involved in using a neural network for multisource classification: the training stage, in which the internal weights are adjusted, and the classifying stage. Typically, the back-propagation algorithm trains the network until some target minimal error is achieved between the desired and actual output values of the network. Once the training is complete, the network is used as a feedforward structure to produce a classification for the entire data set (Paola and Schowengerdt 1995). A neural network consists of a number of interconnected nodes, where each node is a simple processing element that responds to the weighted inputs it receives from other nodes. The arrangement of the nodes is referred to as the ‘‘network architecture’’ (Fig. 5). The receiving node sums the weighted signals from all of the nodes connected to it in the preceding layer. Formally, the input that a single node receives is weighted. The network used in this study consisted of three layers. The first was the input layer, where the nodes were the elements of a feature vector; the second was the internal or ‘‘hidden’’ layers; the third was the output layer that presented the output data. Each node in the hidden layer is interconnected to nodes in both the preceding and following layers by weighted connections (Atkinson and Tatnall 1997). Using the back-propagation training algorithm, the weights of each factor can be determined and may be used for classification of data (input vectors) that the network has not seen before. Zhou (1999) described a method for determining the weights using back-propagation. Procedure The eight physiographic and structural factors were used as the input data. In an ANN, the selection of training sites is important, and the selection of likely and unlikely GSS areas was carefully considered for training in this study. To select the training sites based on scientific and objective criteria, locations considered likely and unlikely to have

Table 2 Weights of hydrogeological factors considered in the GSH analysis based on each case Factor

Case no. 1

2

3

4

5

6

7

8

9

10

Average

SDs

Weight

Slope

0.1066

0.0782

0.1221

0.1103

0.1484

0.1144 0.0867

0.0875

0.1133

0.0776

0.1055

0.0212

1.0555

Depth of drift

0.0999

0.1312

0.1327

0.1149

0.1159

0.1144 0.1398

0.1568

0.1191

0.1347

0.1302

0.0229

1.3026

Distance from drift

0.1041

0.1497

0.1504

0.1052

0.1238

0.1101 0.0875

0.1153

0.1574

0.1458

0.1183

0.0268

1.1839

Depth of groundwater

0.1403

0.1133

0.0929

0.1348

0.1093

0.1547 0.1124

0.1126

0.1222

0.1123

0.1180

0.0208

1.1804

RMR

0.1118

0.0981

0.1222

0.1096

0.0792

0.1067 0.0946

0.0795

0.1052

0.0930

0.0999

0.0139

1.0000

Distance from lineament

0.1926

0.1766

0.1126

0.1192

0.1164

0.1384 0.1669

0.1870

0.1415

0.1963

0.1548

0.0330

1.5491

Geology

0.1110

0.1296

0.1281

0.1671

0.1481

0.1326 0.1513

0.1335

0.1451

0.1189

0.1375

0.0157

1.3756

Land use

0.1336

0.1233

0.1389

0.1387

0.1591

0.1286 0.1608

0.1279

0.0961

0.1214

0.1359

0.0192

1.3604

123

356

achieved, then the maximum number of iterations was terminated at 5,000 epochs. An epoch means the entire training set to the neural network. The maximum RMSE value when the latter case occurred was 0.214. The final weights between layers acquired during training of the neural network, and the contribution or importance of each of the eight factors, were used to predict GSS. Finally, the weights were applied to the entire study area, and GSS maps were created for each training case.

Results Weight Determination and GSS Mapping The final weights between layers acquired during training of the neural network and the contribution or importance of each of the eight factors used to predict GSS are listed in Table 2. The results were not the same because the initial weights were assigned random values. In this study, the calculations were repeated ten times to allow the results to achieve similar values. The SD of the results ranged from 0.014 to 0.033; therefore, random sampling did not have a large effect on the results. For easy interpretation, the average values were calculated, and these values were divided by the average of the weights of the factor having the minimum value. Among the weights, ‘‘distance from lineament’’ had the highest value (1.5477) and ‘‘RMR’’ had the lowest (1.000). The weights were applied to the entire study area, and the GSS maps were created. The GSS map for case no. 1, which had the best accuracy after validation, is shown in

Fig. 6 GSS map using ANN

123

Environmental Management (2012) 49:347–358

Fig. 6. The GSS values were classified by equal areas and grouped into five classes (% of area) of GSS rank for easy visual interpretation: very high (5%), high (5%), medium (10%), low (20%), and low (60%). The minimum and maximum values were 0.008 and 0.952, respectively. The mean and SD were 0.205 and 0.221, respectively. Validation The susceptibility maps constructed using ANN were validated with information on existing ground subsidence. For validation, predictions made using the ANN method were compared with expected results based on knowledge of the factors. For this procedure, rate curves were constructed. To obtain the rate curves, the calculated GSS values of all grids in the study area were sorted in descending order. The ordered grid values were then divided into 100 classes in accumulated 1% intervals. The rate curves explain how well the method and factors predict ground subsidence. Then AUC was calculated to compare the results. In AUC analysis, a total area of 1 denotes perfect prediction accuracy for all cases. The AUC method can be used to assess prediction accuracy qualitatively. The percentages of validated results appear as a line in Fig. 7. For example, for case no. 1, which showed the highest accuracy, 10% of the study area having a greater GSS could explain 90% of all ground subsidence. In addition, 20% of the study area where the GSS value had a greater rank could explain 99% of ground subsidence (Fig. 7). AUC values for the ANNproduced GSS maps were between 0.9484 and 0.9598, meaning that the GSS maps had accuracies between 94.84 and 95.98%.

Environmental Management (2012) 49:347–358

Fig. 7 Cumulative frequency diagram showing ground-subsidence hazard index rank (x-axis) occurring in cumulative percent of groundsubsidence area (y-axis)

Conclusion and Discussion It may take a long time from the beginning of ground subsidence in underground cavities until visible damages occur at the surface. When an underground mine is abandoned, ground subsidence develops from the mine cavity roof to the ground surface (Kwon and others 2000). Various factors, such as construction of underground facilities, abandoned coal mines, soft soil in a landfill area, and corrosion of limestone, can generate ground subsidence. Likewise, ground subsidence includes various factors and complex relations among those factors. Therefore, it is necessary to have an overall and systematic analysis method for understanding the effects of each factor and the interactions among factors. In this study, we used GIS techniques to study the prediction and management of ground subsidence in abandoned coal mines. An ANN model was applied to assess and predict GSH in Jeong-am, South Korea, a region where ground subsidence is expected to continue in the future. The influences of factors that are expected to affect ground subsidence were quantitatively analyzed. Maps of GSS were made using an ANN and repeated ten times. Training sites were extracted from ground-subsidence areas. Validation showed between 94.84 and 95.98% prediction accuracy (average of 95.41%). The accuracies were similar and satisfactory considering the map scale and input data accuracy. Relative environmental factors played important roles in producing the final map products. Given the relative weight of various physiographic and structural factors (Table 2), it

357

was not surprising that RMR had little influence on the final model results considering the paucity of data and the narrow value variation range for those few data that were extrapolated. However, this does not mean that RMR is an unimportant factor in defining subsidence hazard. Rather, our available data were insufficient to support any other conclusion. The spatial distribution of the lineament data extended across the entire area, and thus these data are expected to have had a greater influence on model results. The primary value of our results is that, even with some incomplete data sets and broad assumptions, the method proved to be a robust and useful tool for estimating and mapping subsidence hazard. In this study, using the ANN, the relative importance and weights of factors were calculated. ‘‘Distance from fault’’ showed the highest value of 1.5477, followed by ‘‘geology,’’ which had a value of 1.3654. ‘‘RMR’’ showed the lowest value at 1.000, and ‘‘slope’’ had a value of 1.0452. These results indicate that ‘‘distance from lineament’’ was the most important factor, being 1.5 times more important than ‘‘RMR’’ in GSS mapping. Factors related with coal mine, such as ‘‘depth of pit’’ and ‘‘distance from pit,’’ were of medium importance. The determined weights indicate that geological factors, such as geology and lineament, are important for ground subsidence compared with other factors, such as distance from coal mine and depth of pit. Regarding topographic factors, ‘‘slope’’ was not as important as ‘‘ground subsidence.’’ Because RMR data have limited availability and low accuracy, RMR showed the lowest weight. This study has showed factors involved in ground subsidence, and the method and findings can be applied to GSS mapping in other regions. Moreover, the GSS map produced can be used to mitigate hazards to people and facilities and as basic data for establishing plans to prevent ground hazards, such as in locating monitoring and facility sites. However, to generalize factors of ground subsidence, more case studies and models are needed. Acknowledgments The authors thank the Coal Industry Promotion Board for providing entire investigation reports and the basic GIS database. This research was supported by the Basic Research Project of the Korea Institute of Geoscience and Mineral Resources funded by the Ministry of Knowledge and Economy of Korea.

References Adinarayana J, Krishna NR (1996) Integration of multi-seasonal remotely-sensed images for improved landuse classification of a hilly watershed using geographical information systems. International Journal of Remote Sensing 17:1679–1688 Ambrozˇicˇ T, Turk G (2003) Prediction of subsidence due to underground mining by artificial neural networks. Computers & Geosciences 29(5):627–637 Atkinson PM, Tatnall ARL (1997) Neural networks in remote sensing. International Journal of Remote Sensing 18:699–709

123

358 Bieniawski ZT (1989) Engineering rock mass classifications: a complete manual for engineers and geologists in mining, civil, and petroleum engineering. Wiley-Interscience, London, pp 40–47 Choi JK, Kim KD, Lee S, Won JS (2010) Application of a fuzzy operator to susceptibility estimations of coal mine subsidence in Taebaek City, Korea. Environmental Earth Science 59(5):1009–1022 Coal Industry Promotion Board (1997) A study on the mechanism of subsidence over abandoned mine area and the Construction method of subsidence prevention. Coal Industry Promotion Board, Seoul, pp 97–106 Esaki T, Djamaluddin I, Mitani Y (2008) A GIS-based prediction method to evaluate subsidence-induced damage from coal mining beneath a reservoir, Kyushu, Japan. Quarterly Journal of Engineering Geology and Hydrogeology 41(3):381–392 Garrett JH (1994) Where and why artificial neural networks are applicable in civil engineering. Journal of Computing in Civil Engineering 8(2):129–130 Geological Society of Korea (1999) The geology of Korea. Sigma Press, Seoul, p 802 Hines JW (1997) Fuzzy and neural approaches in engineering. Wiley, New York, p 209 Kim KD, Lee S, Oh HJ, Choi JK, Won JS (2006) Assessment of ground subsidence hazard near an abandoned underground coal mine using GIS. Environmental Geology 50(8):1183–1191 Kim KD, Lee S, Oh HJ (2009) Prediction of ground subsidence in Samcheok City, Korea using artificial neural networks and GIS. Environmental Geology 58(1):61–70 Kwon K, Yu MH, Park HD (2000) A study on the application of GIS for analysis of subsidence hazard. Economic and Environmental Geology 33(6):557–563 Lee S, Dan NT (2005) Probabilistic landslide susceptibility mapping in the Lai Chau province of Vietnam: focus on the relationship between tectonic fractures and landslides. Environmental Geology 48:778–787 Lee S, Oh HJ, Kim KD (2010) Statistical spatial modeling of ground subsidence hazard near an abandoned underground coal mine. Disaster Advances 3(1):11–23

123

Environmental Management (2012) 49:347–358 Mancini F, Stecchi F, Gabbianelli G (2009) GIS-based assessment of risk due to salt mining activities at Tuzla (Bosnia and Herzegovina). Engineering Geology 109(3–4):170–182 Mine Reclamation Corporation (2007). http://www.mireco.or.kr. Accessed 10 August 2007 Ministry of Knowledge Economy (2007) Available at: http://www.mocie.go.kr. Accessed 16 July 2007 Oh HJ, Lee S (2010) Assessment of ground subsidence using GIS and the weights-of-evidence model. Engineering Geology 115:26–48 Oh HJ, Lee S (2011) Integration of ground subsidence hazard maps of abandoned coal mines in Samcheok, Korea. International Journal of Coal Geology 86(1):58–72 Oh HJ, Ahn SC, Choi JK, Lee S (2011) Sensitivity analysis for the GIS-based mapping of the ground subsidence hazard near abandoned underground coal mines. Environmental Earth Science 64:347–358 Paola JD, Schowengerdt RA (1995) A review and analysis of backpropagation neural networks for classification of remotely sensed multi-spectral imagery. International Journal of Remote Sensing 16:3033–3058 Quanyuan W, Jiewu P, Shanzhong Q, Yiping L, Congcong H, Tingxiang L et al (2009) Impacts of coal mining subsidence on the surface landscape in Longkou city, Shandong Province of China. Environmental Earth Science 59(4):783–791 Turer D, Nefeslioglu HA, Zorlu K, Gokceoglu C (2008) Assessment of geo-environmental problems of the Zonguldak province (NW Turkey). Environmental Geology 55(5):1001–1014 Waltham AC (1989) Ground subsidence. Blackie & Son Ltd, New York, pp 49–97 Zahiri H, Palamara DR, Flentje P, Brassington GM, Baafi E (2006) A GIS-based weights-of-evidence model for mapping cliff instabilities associated with mine subsidence. Environmental Geology 51(3):377–386 Zhou W (1999) Verification of the nonparametric characteristics of backpropagation neural networks for image classification. IEEE Transactions on Geoscience and Remote Sensing 37(2):771–779