Scholarlink

<p>Landslides represent a major hazard in the tectonically active Himalayan region, where complex terrain, hydrological perturbations, and expanding infrastructure strongly influence slope stability. Conventional landslide susceptibility approaches that rely on globally uniform conditioning factors often fail to capture localized geomorphic and structural controls, leading to reduced reliability in highly heterogeneous Himalayan terrain. To address this limitation, this study conducts landslide susceptibility mapping (LSM) for the Bhagirathi catchment by evaluating the effectiveness of terrain-adaptive conditioning factor selection for hazard characterization. Analysis of a quality-controlled landslide inventory derived from Geological Survey of India records reveals distinct and spatially coherent landslide susceptibility patterns across the basin. The results reveal clear differences in predictive performance among models, with Logistic Regression exhibiting the lowest discriminatory capability (AUC = 0.88), while Gradient Boosting achieved the highest accuracy (AUC = 0.93; Kappa = 0.69), providing the most reliable separation between stable and unstable slopes. The susceptibility maps indicate that approximately 14% of the Bhagirathi catchment falls within high to very high susceptibility classes, predominantly concentrated along river corridors, reservoir-affected slopes, and structurally weak lithological units. These high-risk zones spatially coincide with areas of elevated infrastructure density and socio-economic exposure, underscoring their importance for hazard zonation and risk-informed planning. Overall, the findings demonstrate that incorporating region-specific, adaptive conditioning factors substantially enhances the reliability, interpretability, and practical utility of landslide susceptibility mapping in the Himalayan context, offering actionable guidance for disaster risk reduction and sustainable land-use planning in fragile mountain environments.</p> Graphical Abstract <p></p> <p>This is a visual summary that presents a regionally calibrated machine learning pipeline for landslide susceptibility mapping (LSM), designed to overcome limitations of generalised approaches. The workflow begins with the preparation of conditioning factors derived from DEM derivatives and ancillary datasets. These are grouped into region-dependent variables (e.g., lithology, soil moisture, drainage density, geomorphology, precipitation anomaly, and proximity to rivers, roads, and faults) and generalised topographic parameters (e.g., slope, curvature, aspect, elevation, and land use/land cover). Landslide and non-landslide points from the Bhukosh database form the dependent variable for model calibration. A key novelty lies in the predictor dependency analysis, where the information gain ratio (IGR) ranks features while identifying redundant or collinear predictors. Low-contribution collinear factors are removed, while non-collinear variables are retained, producing a more parsimonious and interpretable dataset. The refined data is divided into training and testing subsets (70:20) for robust validation. The machine learning pipeline is implemented in scikit-learn with data handling in pandas. Algorithms such as logistic regression, random forest, gradient boosting, extreme gradient boosting, k-nearest neighbours, and support vector machines are calibrated using both generalised and region-specific features. The resulting susceptibility maps classify terrain into very low to very high-risk zones, capturing regional variability and predictive reliability. This study is particularly important for the Bhagirathi catchment in the Lower Himalayas, a landslide-prone region with fragile lithology, steep slopes, intense rainfall, and active tectonics. By integrating regional calibration and predictor dependency analysis, the framework improves accuracy, reduces bias, and enhances interpretability compared to conventional models. The maps provide actionable insights for disaster managers, engineers, and policymakers, while the pipeline is transferable to other Himalayan catchments, offering a scalable tool for landslide hazard assessment in data-scarce, geologically complex terrains.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Beyond Generalisation: Regionally Calibrated Feature Selection with Machine Learning for Data-driven Landslide Hazard Assessment in the Himalayas

  • 期刊论文

  • 发表日期：

    2026年2月20日

查看全文
  • Madhukar Dwivedi,
  • Srikrishnan Sivasubramanian,
  • Kamal Jain,
  • Ali P Yunus

摘要

Landslides represent a major hazard in the tectonically active Himalayan region, where complex terrain, hydrological perturbations, and expanding infrastructure strongly influence slope stability. Conventional landslide susceptibility approaches that rely on globally uniform conditioning factors often fail to capture localized geomorphic and structural controls, leading to reduced reliability in highly heterogeneous Himalayan terrain. To address this limitation, this study conducts landslide susceptibility mapping (LSM) for the Bhagirathi catchment by evaluating the effectiveness of terrain-adaptive conditioning factor selection for hazard characterization. Analysis of a quality-controlled landslide inventory derived from Geological Survey of India records reveals distinct and spatially coherent landslide susceptibility patterns across the basin. The results reveal clear differences in predictive performance among models, with Logistic Regression exhibiting the lowest discriminatory capability (AUC = 0.88), while Gradient Boosting achieved the highest accuracy (AUC = 0.93; Kappa = 0.69), providing the most reliable separation between stable and unstable slopes. The susceptibility maps indicate that approximately 14% of the Bhagirathi catchment falls within high to very high susceptibility classes, predominantly concentrated along river corridors, reservoir-affected slopes, and structurally weak lithological units. These high-risk zones spatially coincide with areas of elevated infrastructure density and socio-economic exposure, underscoring their importance for hazard zonation and risk-informed planning. Overall, the findings demonstrate that incorporating region-specific, adaptive conditioning factors substantially enhances the reliability, interpretability, and practical utility of landslide susceptibility mapping in the Himalayan context, offering actionable guidance for disaster risk reduction and sustainable land-use planning in fragile mountain environments.

Graphical Abstract

This is a visual summary that presents a regionally calibrated machine learning pipeline for landslide susceptibility mapping (LSM), designed to overcome limitations of generalised approaches. The workflow begins with the preparation of conditioning factors derived from DEM derivatives and ancillary datasets. These are grouped into region-dependent variables (e.g., lithology, soil moisture, drainage density, geomorphology, precipitation anomaly, and proximity to rivers, roads, and faults) and generalised topographic parameters (e.g., slope, curvature, aspect, elevation, and land use/land cover). Landslide and non-landslide points from the Bhukosh database form the dependent variable for model calibration. A key novelty lies in the predictor dependency analysis, where the information gain ratio (IGR) ranks features while identifying redundant or collinear predictors. Low-contribution collinear factors are removed, while non-collinear variables are retained, producing a more parsimonious and interpretable dataset. The refined data is divided into training and testing subsets (70:20) for robust validation. The machine learning pipeline is implemented in scikit-learn with data handling in pandas. Algorithms such as logistic regression, random forest, gradient boosting, extreme gradient boosting, k-nearest neighbours, and support vector machines are calibrated using both generalised and region-specific features. The resulting susceptibility maps classify terrain into very low to very high-risk zones, capturing regional variability and predictive reliability. This study is particularly important for the Bhagirathi catchment in the Lower Himalayas, a landslide-prone region with fragile lithology, steep slopes, intense rainfall, and active tectonics. By integrating regional calibration and predictor dependency analysis, the framework improves accuracy, reduces bias, and enhances interpretability compared to conventional models. The maps provide actionable insights for disaster managers, engineers, and policymakers, while the pipeline is transferable to other Himalayan catchments, offering a scalable tool for landslide hazard assessment in data-scarce, geologically complex terrains.