Scholarlink

<p>The NH-44 route connecting Udhampur and Chenani in Jammu and Kashmir, India, is characterized as a severely landslide-prone corridor in the Himalayas. A multi-temporal evaluation (2013–2024) was performed with inventories that included 28, 52, 73, 85, 92, and 83 landslides for the years 2013, 2015, 2017, 2019, 2021, and 2024, respectively. The frequency of landslides reached a maximum of 92 occurrences in 2021, thereafter decreasing to 83 occurrences in 2024. The impacted area expanded steadily from 0.065&#xa0;km² in 2013 to 0.311&#xa0;km² in 2021 and 0.321&#xa0;km² in 2024, signifying a transition towards greater failures and ongoing slope destabilization due to anthropogenic influences. Year-specific susceptibility models were constructed utilizing Frequency Ratio (FR), Random Forest (RF), and Support Vector Machine (SVM) with a uniform set of conditioning parameters. The evaluation of model performance was conducted by ROC–AUC, Precision–Recall analysis, and other classification measures, encompassing Accuracy, Precision, Recall, Specificity, F1-score, Cohen’s Kappa, and True Skill Statistic (TSS). Robustness was evaluated using time cross-validation using the 2024 inventory and spatial validation within a 100&#xa0;m road corridor buffer. All models attained elevated ROC–AUC values (&gt; 0.85), with RF and SVM demonstrating exceptional classification efficacy. Nonetheless, polygon- and frequency-based validation demonstrates that FR encompasses a greater percentage of landslides within the high and very high susceptibility categories. In 2024, over 90% of the landslide area is situated within H–VH zones in FR, in contrast to approximately 60–75% in RF and SVM. Field validation further substantiates the enhanced spatial concordance of FR outputs with the documented extents of landslides along road-cut slopes. Proximity to roads, slope, and temporal land use/land cover (LULC) are recognized as primary determinants, with a growing anthropogenic impact post-2015. Findings indicate that dependence solely on ROC–AUC may inflate the assessment of predictive efficacy. Multi-criteria validation is crucial for a rigorous susceptibility evaluation. The FR model is more efficient for localized, road-corridor-scale applications, while machine learning (ML) approaches are more appropriate for regional-scale generalization.</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

Multi-Model Back Analysis for Predicting Landslide Occurrences Under Extreme Anthropogenic Regime in the Indian Himalaya

  • 期刊论文

  • 发表日期：

    2026年5月4日

查看全文
  • Imran Khan,
  • Ashutosh Kainthola,
  • Harish Bahuguna

摘要

The NH-44 route connecting Udhampur and Chenani in Jammu and Kashmir, India, is characterized as a severely landslide-prone corridor in the Himalayas. A multi-temporal evaluation (2013–2024) was performed with inventories that included 28, 52, 73, 85, 92, and 83 landslides for the years 2013, 2015, 2017, 2019, 2021, and 2024, respectively. The frequency of landslides reached a maximum of 92 occurrences in 2021, thereafter decreasing to 83 occurrences in 2024. The impacted area expanded steadily from 0.065 km² in 2013 to 0.311 km² in 2021 and 0.321 km² in 2024, signifying a transition towards greater failures and ongoing slope destabilization due to anthropogenic influences. Year-specific susceptibility models were constructed utilizing Frequency Ratio (FR), Random Forest (RF), and Support Vector Machine (SVM) with a uniform set of conditioning parameters. The evaluation of model performance was conducted by ROC–AUC, Precision–Recall analysis, and other classification measures, encompassing Accuracy, Precision, Recall, Specificity, F1-score, Cohen’s Kappa, and True Skill Statistic (TSS). Robustness was evaluated using time cross-validation using the 2024 inventory and spatial validation within a 100 m road corridor buffer. All models attained elevated ROC–AUC values (> 0.85), with RF and SVM demonstrating exceptional classification efficacy. Nonetheless, polygon- and frequency-based validation demonstrates that FR encompasses a greater percentage of landslides within the high and very high susceptibility categories. In 2024, over 90% of the landslide area is situated within H–VH zones in FR, in contrast to approximately 60–75% in RF and SVM. Field validation further substantiates the enhanced spatial concordance of FR outputs with the documented extents of landslides along road-cut slopes. Proximity to roads, slope, and temporal land use/land cover (LULC) are recognized as primary determinants, with a growing anthropogenic impact post-2015. Findings indicate that dependence solely on ROC–AUC may inflate the assessment of predictive efficacy. Multi-criteria validation is crucial for a rigorous susceptibility evaluation. The FR model is more efficient for localized, road-corridor-scale applications, while machine learning (ML) approaches are more appropriate for regional-scale generalization.