Landslide Hazard Zonation Using Maximum Entropy (MaxEnt) and Remote Sensing Data in the Shahid Abbaspour Dam Catchment Area
Document Type : Original Article
Authors
1
PhD student, Department of Physical Geography, Faculty of Geographical Sciences and Planning, University of Isfahan, Isfahan, Iran.
2
university of Isfahan
3
Associate Professor, Department of Marine Geology, Faculty of Marine Natural Resources, Khorramshahr University of Marine Sciences and Technology, Khorramshahr, Iran.
10.22034/gmpj.2026.573236.1600
Abstract
Extended Abstract
Introduction
Landslides are among the most devastating and costliest natural hazards worldwide, posing significant challenges to the sustainable management of mountainous and steep-slope regions. Iran, due to its diverse geological and climatic settings, is highly susceptible to various natural disasters, with landslides being particularly prevalent in the Zagros folded belt. The northeastern part of Khuzestan province, encompassing the Shahid Abbaspour Dam basin and its sub-basins (Northern and Southern Deh-Sheikh), experiences frequent landslide events due to a combination of active tectonics, erodible lithological formations, intense seasonal rainfall, and anthropogenic interventions such as road construction and land-use changes. Although numerous studies have employed machine learning algorithms for landslide susceptibility mapping, a comprehensive comparative assessment of the controlling factors across multiple watersheds with distinct environmental settings remains lacking. This study, therefore, aims to: (1) evaluate the performance of the Maximum Entropy (MaxEnt) algorithm, (2) identify and prioritize the most influential environmental factors, (3) produce spatially explicit landslide hazard zonation maps, and (4) compare the driving mechanisms among the three watersheds, integrating field surveys, remote sensing data (Sentinel-2, ASTER GDEM), and GIS-based geospatial analyses.
Methodology
A total of 129 landslide occurrence points were compiled from geological survey databases, extensive field campaigns, and high-resolution Google Earth imagery interpretation. Fifteen conditioning factors were selected based on a systematic literature review and local geo-environmental conditions, including topographic parameters (elevation, slope, aspect, plan curvature), geological settings (lithology, distance to faults), hydrological indices (Topographic Wetness Index, Stream Power Index, distance to rivers), anthropogenic factors (distance to roads, land use/land cover), climatic variable (mean annual precipitation), soil texture, and vegetation cover (NDVI). All layers were resampled to a uniform spatial resolution of 30 m and standardized to a common coordinate system (UTM Zone 39N). Multicollinearity among predictors was assessed using the Variance Inflation Factor (VIF), and variables with VIF > 10 were excluded. The MaxEnt model was calibrated using 70% of the landslide points for training and 30% for testing, with 10-fold cross-validation to ensure robustness. Model performance was evaluated using the Receiver Operating Characteristic (ROC) curve and the Area Under the Curve (AUC), supplemented by the True Skill Statistic (TSS) and Kappa coefficient. Variable importance was assessed through percent contribution and the Jackknife test, while response curves were employed to elucidate the nature of factor-landslide relationships. The final susceptibility maps were classified into five hazard classes (very low to very high) using the Natural Breaks (Jenks) optimization method.
Results
The MaxEnt model demonstrated excellent predictive performance across all three watersheds, with AUC values exceeding 0.90 for both training and testing datasets (Shahid Abbaspour: 0.908 and 0.906; Northern Deh-Sheikh: 0.970 and 0.908; Southern Deh-Sheikh: 0.929 and 0.978, respectively). These high AUC values, along with the close agreement between training and testing results, indicate robust model generalization and negligible overfitting. Jackknife tests revealed that all variables contributed positively to the model, but their relative importance varied markedly among watersheds. In the Shahid Abbaspour basin, precipitation (28.41%), lithology (20.99%), distance to roads (13.93%), and distance to faults (9.37%) were the dominant factors, with precipitation showing a clear threshold behavior at ~2600 mm/year. In Northern Deh-Sheikh, land use/land cover (14.07%), distance to roads (13.88%), distance to rivers (12.39%), and Stream Power Index (11.09%) were the primary controls, highlighting the overriding role of anthropogenic and hydrological drivers. In Southern Deh-Sheikh, lithology (24.99%), distance to roads (19.93%), distance to rivers (10.40%), and precipitation (10.39%) were most influential, with Quaternary deposits and Gachsaran Formation being the most susceptible lithological units. The hazard maps revealed distinct spatial patterns: the Shahid Abbaspour basin exhibited the largest area under high-hazard class (78.08 km², 30.19%), Northern Deh-Sheikh was predominantly classified as moderate hazard (79.55 km², 40.21%), while Southern Deh-Sheikh showed the most critical condition with the largest very-high-hazard zone (79.84 km², 33.21%). The combined high and very-high hazard classes covered over 53%, 27%, and 64% of the total area in the three basins, respectively.
Discussion and Conclusion
The results highlight the critical role of local environmental heterogeneity in landslide susceptibility. Inter-basin differences in factor importance—precipitation and lithology in Shahid Abbaspour and Southern Deh‑Sheikh versus anthropogenic/hydrological factors in Northern Deh‑Sheikh—reflect the complex interaction between natural predisposing factors and human triggers. The high susceptibility of the Gachsaran Formation (gypsum‑ and salt‑bearing evaporites) and Quaternary unconsolidated deposits explains elevated hazard levels in these units. A precipitation threshold of ~2600–3200 mm/year suggests saturation‑driven failure mechanisms predominate in high‑rainfall catchments. The consistent influence of roads across all basins underscores the role of anthropogenic slope modifications in exacerbating susceptibility. Excellent model performance (AUC, TSS, Kappa) confirms MaxEnt as a robust tool for landslide assessment in data‑scarce regions. However, limitations include imbalanced landslide sampling, coarse geological maps (1:100,000), and sole reliance on AUC, which may not fully capture performance in imbalanced datasets. Future work should employ data balancing techniques, integrate physically‑based models (e.g., TRIGRS) with statistical approaches, incorporate higher‑resolution geological/geotechnical data, and assess climate change impacts on rainfall‑induced slope instability. The susceptibility maps provide a scientific basis for prioritizing conservation, guiding urban planning, and developing region‑specific mitigation strategies in the Zagros and similar mountainous regions worldwide.
Introduction
Landslides are among the most devastating and costliest natural hazards worldwide, posing significant challenges to the sustainable management of mountainous and steep-slope regions. Iran, due to its diverse geological and climatic settings, is highly susceptible to various natural disasters, with landslides being particularly prevalent in the Zagros folded belt. The northeastern part of Khuzestan province, encompassing the Shahid Abbaspour Dam basin and its sub-basins (Northern and Southern Deh-Sheikh), experiences frequent landslide events due to a combination of active tectonics, erodible lithological formations, intense seasonal rainfall, and anthropogenic interventions such as road construction and land-use changes. Although numerous studies have employed machine learning algorithms for landslide susceptibility mapping, a comprehensive comparative assessment of the controlling factors across multiple watersheds with distinct environmental settings remains lacking. This study, therefore, aims to: (1) evaluate the performance of the Maximum Entropy (MaxEnt) algorithm, (2) identify and prioritize the most influential environmental factors, (3) produce spatially explicit landslide hazard zonation maps, and (4) compare the driving mechanisms among the three watersheds, integrating field surveys, remote sensing data (Sentinel-2, ASTER GDEM), and GIS-based geospatial analyses.
Methodology
A total of 129 landslide occurrence points were compiled from geological survey databases, extensive field campaigns, and high-resolution Google Earth imagery interpretation. Fifteen conditioning factors were selected based on a systematic literature review and local geo-environmental conditions, including topographic parameters (elevation, slope, aspect, plan curvature), geological settings (lithology, distance to faults), hydrological indices (Topographic Wetness Index, Stream Power Index, distance to rivers), anthropogenic factors (distance to roads, land use/land cover), climatic variable (mean annual precipitation), soil texture, and vegetation cover (NDVI). All layers were resampled to a uniform spatial resolution of 30 m and standardized to a common coordinate system (UTM Zone 39N). Multicollinearity among predictors was assessed using the Variance Inflation Factor (VIF), and variables with VIF > 10 were excluded. The MaxEnt model was calibrated using 70% of the landslide points for training and 30% for testing, with 10-fold cross-validation to ensure robustness. Model performance was evaluated using the Receiver Operating Characteristic (ROC) curve and the Area Under the Curve (AUC), supplemented by the True Skill Statistic (TSS) and Kappa coefficient. Variable importance was assessed through percent contribution and the Jackknife test, while response curves were employed to elucidate the nature of factor-landslide relationships. The final susceptibility maps were classified into five hazard classes (very low to very high) using the Natural Breaks (Jenks) optimization method.
Results
The MaxEnt model demonstrated excellent predictive performance across all three watersheds, with AUC values exceeding 0.90 for both training and testing datasets (Shahid Abbaspour: 0.908 and 0.906; Northern Deh-Sheikh: 0.970 and 0.908; Southern Deh-Sheikh: 0.929 and 0.978, respectively). These high AUC values, along with the close agreement between training and testing results, indicate robust model generalization and negligible overfitting. Jackknife tests revealed that all variables contributed positively to the model, but their relative importance varied markedly among watersheds. In the Shahid Abbaspour basin, precipitation (28.41%), lithology (20.99%), distance to roads (13.93%), and distance to faults (9.37%) were the dominant factors, with precipitation showing a clear threshold behavior at ~2600 mm/year. In Northern Deh-Sheikh, land use/land cover (14.07%), distance to roads (13.88%), distance to rivers (12.39%), and Stream Power Index (11.09%) were the primary controls, highlighting the overriding role of anthropogenic and hydrological drivers. In Southern Deh-Sheikh, lithology (24.99%), distance to roads (19.93%), distance to rivers (10.40%), and precipitation (10.39%) were most influential, with Quaternary deposits and Gachsaran Formation being the most susceptible lithological units. The hazard maps revealed distinct spatial patterns: the Shahid Abbaspour basin exhibited the largest area under high-hazard class (78.08 km², 30.19%), Northern Deh-Sheikh was predominantly classified as moderate hazard (79.55 km², 40.21%), while Southern Deh-Sheikh showed the most critical condition with the largest very-high-hazard zone (79.84 km², 33.21%). The combined high and very-high hazard classes covered over 53%, 27%, and 64% of the total area in the three basins, respectively.
Discussion and Conclusion
The results highlight the critical role of local environmental heterogeneity in landslide susceptibility. Inter-basin differences in factor importance—precipitation and lithology in Shahid Abbaspour and Southern Deh‑Sheikh versus anthropogenic/hydrological factors in Northern Deh‑Sheikh—reflect the complex interaction between natural predisposing factors and human triggers. The high susceptibility of the Gachsaran Formation (gypsum‑ and salt‑bearing evaporites) and Quaternary unconsolidated deposits explains elevated hazard levels in these units. A precipitation threshold of ~2600–3200 mm/year suggests saturation‑driven failure mechanisms predominate in high‑rainfall catchments. The consistent influence of roads across all basins underscores the role of anthropogenic slope modifications in exacerbating susceptibility. Excellent model performance (AUC, TSS, Kappa) confirms MaxEnt as a robust tool for landslide assessment in data‑scarce regions. However, limitations include imbalanced landslide sampling, coarse geological maps (1:100,000), and sole reliance on AUC, which may not fully capture performance in imbalanced datasets. Future work should employ data balancing techniques, integrate physically‑based models (e.g., TRIGRS) with statistical approaches, incorporate higher‑resolution geological/geotechnical data, and assess climate change impacts on rainfall‑induced slope instability. The susceptibility maps provide a scientific basis for prioritizing conservation, guiding urban planning, and developing region‑specific mitigation strategies in the Zagros and similar mountainous regions worldwide.
Keywords
