Abstract Landslides are a significant geohazard in the northeastern states of India due to intense precipitation and unstable geo-environmental settings. This study uses novel machine learning (ML) approaches and twenty-six landslide causative factors (LCFs) to develop landslide susceptibility models for Nagaland, India. We employed six state-of-the-art ML models: linear discriminant analysis, bagged CART, k-nearest neighbours, random forest, support vector machine, and artificial neural networks. To determine the most important LCFs, we evaluated the models’ performance using progressively larger subsets of LCFs (the top 5, 10, 15, 20, and 23), which were ranked based on their relative importance computed using the gain ratio feature selection method. The ML models were assessed with several metrics, including the area under the curve (AUC), sensitivity, and specificity, using 10-fold cross-validation and separate test datasets. Ensemble models, such as random forest and bagged CART, demonstrated significantly higher performance. Using only the top five LCFs (soil moisture, average annual precipitation, distance to major roads, geology, and distance to earthquake epicenters), these models achieved an AUC of 94–97%. The resulting landslide susceptibility maps revealed that ~20% of Nagaland falls within the high to very high susceptibility categories. Most of these regions are near major settlements, posing a significant threat to the population. The susceptibility maps produced in this study should act as vital tools for developing effective landslide mitigation strategies in Nagaland. Research highlights <OrderedList> <ListItem> <ItemNumber>1.</ItemNumber> <ItemContent> Developed statewide landslide susceptibility maps for Nagaland, India, using state-of-the-art machine learning models. </ItemContent> </ListItem> <ListItem> <ItemNumber>2.</ItemNumber> <ItemContent> Soil moisture, precipitation, distance to roads, geology, and distance to earthquake epicenters are the major causative factors of landslides in Nagaland. </ItemContent> </ListItem> <ListItem> <ItemNumber>3.</ItemNumber> <ItemContent> The random forest model achieved the highest accuracy, with an AUC of 94–97% and sensitivity and specificity ranging from 88–97%. </ItemContent> </ListItem> <ListItem> <ItemNumber>4.</ItemNumber> <ItemContent> About 20% of Nagaland is highly susceptible to landslides, with high-risk zones concentrated around major settlements. </ItemContent> </ListItem> </OrderedList> Graphical abstract

State-level landslide susceptibility mapping using machine learning models: A case study of Nagaland, India

期刊论文
发表日期：

2026年1月24日

查看全文

Sourav Chetia,
Chandan Kumar,
Sujata Medhi,
Anup Saikia

摘要

Abstract

Landslides are a significant geohazard in the northeastern states of India due to intense precipitation and unstable geo-environmental settings. This study uses novel machine learning (ML) approaches and twenty-six landslide causative factors (LCFs) to develop landslide susceptibility models for Nagaland, India. We employed six state-of-the-art ML models: linear discriminant analysis, bagged CART, k-nearest neighbours, random forest, support vector machine, and artificial neural networks. To determine the most important LCFs, we evaluated the models’ performance using progressively larger subsets of LCFs (the top 5, 10, 15, 20, and 23), which were ranked based on their relative importance computed using the gain ratio feature selection method. The ML models were assessed with several metrics, including the area under the curve (AUC), sensitivity, and specificity, using 10-fold cross-validation and separate test datasets. Ensemble models, such as random forest and bagged CART, demonstrated significantly higher performance. Using only the top five LCFs (soil moisture, average annual precipitation, distance to major roads, geology, and distance to earthquake epicenters), these models achieved an AUC of 94–97%. The resulting landslide susceptibility maps revealed that ~20% of Nagaland falls within the high to very high susceptibility categories. Most of these regions are near major settlements, posing a significant threat to the population. The susceptibility maps produced in this study should act as vital tools for developing effective landslide mitigation strategies in Nagaland.

Research highlights

Developed statewide landslide susceptibility maps for Nagaland, India, using state-of-the-art machine learning models.

Soil moisture, precipitation, distance to roads, geology, and distance to earthquake epicenters are the major causative factors of landslides in Nagaland.

The random forest model achieved the highest accuracy, with an AUC of 94–97% and sensitivity and specificity ranging from 88–97%.

About 20% of Nagaland is highly susceptible to landslides, with high-risk zones concentrated around major settlements.

Graphical abstract