Abstract Landslide susceptibility assessment at the regional scale is an effective tool for sustainable risk management. To enhance the reliability of susceptibility analysis, selecting an optimised dataset for model training/testing is paramount. This study proposes a novel method for landslide susceptibility analysis using fresh landslide inventory, explainable machine learning (ML), and a statistically consistent method that preserves the statistical properties of the input datasets for ML models. The proposed framework is used to generate landslide susceptibility maps along National Highway 707 in Himachal Pradesh, India. In this study, 10 geo-environmental factors and a landslide inventory of the region are partitioned for training and testing the ML models using the statistically consistent method, rather than a simple random sampling method, for preparing landslide susceptibility. Results indicate that among the three ML models of random forest, logistic regression and support vector machine, the random forest model outperforms the other two models with a higher AUC value of 0.932. The trained random forest model is interpreted using Shapley values to alleviate the intrinsic opacity of ML models, i.e., “black box” nature. Finally, the developed landslide susceptibility maps were validated through field observations, contour maps developed using a survey-grade drone, and past studies on the landslide deformation in the study area. This study highlights the importance of improved ML models for generating landslide susceptibility maps, which are useful for engineers in careful planning for infrastructure development in landslide-prone areas. Research highlights <UnorderedList Mark="Bullet"> <ItemContent> Statistically consistent (SC) datasets used for improved landslide susceptibility. </ItemContent> <ItemContent> Random forest trained with SC datasets outperformed other ML models. </ItemContent> <ItemContent> Proposed SHAP-based explainable ML framework for landslide susceptibility. </ItemContent> <ItemContent> Proposed model can identify landslide-prone areas along Highway in Himachal Pradesh, India. </ItemContent> </UnorderedList>

Improved landslide susceptibility mapping framework using explainable machine learning and statistically consistent datasets: a case study for a National Highway in Himachal Pradesh, India

期刊论文
发表日期：

2026年6月22日

查看全文

Sonam Ladol,
Alok Bhardwaj,
Satyendra Mittal,
Mark Brian Jaksa

摘要

Abstract

Landslide susceptibility assessment at the regional scale is an effective tool for sustainable risk management. To enhance the reliability of susceptibility analysis, selecting an optimised dataset for model training/testing is paramount. This study proposes a novel method for landslide susceptibility analysis using fresh landslide inventory, explainable machine learning (ML), and a statistically consistent method that preserves the statistical properties of the input datasets for ML models. The proposed framework is used to generate landslide susceptibility maps along National Highway 707 in Himachal Pradesh, India. In this study, 10 geo-environmental factors and a landslide inventory of the region are partitioned for training and testing the ML models using the statistically consistent method, rather than a simple random sampling method, for preparing landslide susceptibility. Results indicate that among the three ML models of random forest, logistic regression and support vector machine, the random forest model outperforms the other two models with a higher AUC value of 0.932. The trained random forest model is interpreted using Shapley values to alleviate the intrinsic opacity of ML models, i.e., “black box” nature. Finally, the developed landslide susceptibility maps were validated through field observations, contour maps developed using a survey-grade drone, and past studies on the landslide deformation in the study area. This study highlights the importance of improved ML models for generating landslide susceptibility maps, which are useful for engineers in careful planning for infrastructure development in landslide-prone areas.

Research highlights

Statistically consistent (SC) datasets used for improved landslide susceptibility.

Random forest trained with SC datasets outperformed other ML models.

Proposed SHAP-based explainable ML framework for landslide susceptibility.

Proposed model can identify landslide-prone areas along Highway in Himachal Pradesh, India.