Abstract The Darjeeling Himalayas, known for its complex geological setting and history of devastating landslides, require robust and reliable landslide susceptibility mapping for effective disaster risk reduction. This study addresses the challenge of spatio-temporal variability in landslide occurrences by utilizing a comprehensive landslide inventory (2453 events from 1968 to 2023) and a novel ensemble modelling approach. Bivariate statistical analysis (Multiclass Index Overlay) was integrated with multivariate machine learning techniques (logistic regression, linear discriminant analysis, and quadratic discriminant analysis) within a stacking framework. Rigorous validation using a post-2017 landslide dataset demonstrated the ensemble model’s superior performance, achieving an Area Under the Curve (AUC) of 0.81 and a balanced accuracy of 0.74. The ensemble model identified high susceptibility zones, primarily in the northern and eastern parts of the study area, corresponding to regions with steep slopes, high relative relief, and the presence of highly weathered geological formations. The spatial variability between the individual models highlighted the complex interplay of factors influencing landslide occurrence and the importance of considering multiple modelling perspectives. This study underscores the power of ensemble methods in achieving accurate and balanced landslide susceptibility maps, crucial for informed decision-making and effective risk management in this vulnerable region of the Darjeeling Himalayas. Research highlights <UnorderedList Mark="Bullet"> <ItemContent> This study describes a novel ensemble methodology that integrates bivariate statistical analysis, specifically Multiclass Index Overlay, with several multivariate machine learning techniques, including Logistic Regression (LR), Linear Discriminant Analysis (LDA), and Quadratic Discriminant Analysis (QDA). </ItemContent> <ItemContent> A comprehensive landslide inventory comprising 2,453 landslides from 1968 to 2023 was employed. The dataset was temporally partitioned to enable independent validation, with events prior to 2018 used for training and those after 2017 reserved for testing. 3. The ensemble model demonstrated heightened predictive accuracy in comparison to to individual base learners, achieving an Area Under the Curve (AUC) of 0.81 and a balanced accuracy of 0.74. </ItemContent> <ItemContent> Critical high-susceptibility zones were predominantly identified in the northern and eastern regions, which are characterised by steep slopes, significant relative relief, and extensively weathered geological formations. </ItemContent> <ItemContent> The stacking ensemble approach adeptly balanced sensitivity and specificity, resulting in a more reliable and stable susceptibility map to support regional disaster risk management. </ItemContent> </UnorderedList>

Stacking ensemble for improved landslide susceptibility mapping in Darjeeling Himalayas, India

期刊论文
发表日期：

2026年2月20日

查看全文

Sumit Kumar,
Gargi Singh,
Rabisankar Karmakar,
Akshaya Kumar Mishra

摘要

Abstract

The Darjeeling Himalayas, known for its complex geological setting and history of devastating landslides, require robust and reliable landslide susceptibility mapping for effective disaster risk reduction. This study addresses the challenge of spatio-temporal variability in landslide occurrences by utilizing a comprehensive landslide inventory (2453 events from 1968 to 2023) and a novel ensemble modelling approach. Bivariate statistical analysis (Multiclass Index Overlay) was integrated with multivariate machine learning techniques (logistic regression, linear discriminant analysis, and quadratic discriminant analysis) within a stacking framework. Rigorous validation using a post-2017 landslide dataset demonstrated the ensemble model’s superior performance, achieving an Area Under the Curve (AUC) of 0.81 and a balanced accuracy of 0.74. The ensemble model identified high susceptibility zones, primarily in the northern and eastern parts of the study area, corresponding to regions with steep slopes, high relative relief, and the presence of highly weathered geological formations. The spatial variability between the individual models highlighted the complex interplay of factors influencing landslide occurrence and the importance of considering multiple modelling perspectives. This study underscores the power of ensemble methods in achieving accurate and balanced landslide susceptibility maps, crucial for informed decision-making and effective risk management in this vulnerable region of the Darjeeling Himalayas.

Research highlights

This study describes a novel ensemble methodology that integrates bivariate statistical analysis, specifically Multiclass Index Overlay, with several multivariate machine learning techniques, including Logistic Regression (LR), Linear Discriminant Analysis (LDA), and Quadratic Discriminant Analysis (QDA).

A comprehensive landslide inventory comprising 2,453 landslides from 1968 to 2023 was employed. The dataset was temporally partitioned to enable independent validation, with events prior to 2018 used for training and those after 2017 reserved for testing. 3. The ensemble model demonstrated heightened predictive accuracy in comparison to to individual base learners, achieving an Area Under the Curve (AUC) of 0.81 and a balanced accuracy of 0.74.

Critical high-susceptibility zones were predominantly identified in the northern and eastern regions, which are characterised by steep slopes, significant relative relief, and extensively weathered geological formations.

The stacking ensemble approach adeptly balanced sensitivity and specificity, resulting in a more reliable and stable susceptibility map to support regional disaster risk management.