Landslides in the North Eastern Region of India, driven by heavy rainfall, soil composition, topography, and land use, highlight the urgent need for a landslide early warning information system (LEWS) to support preparedness and mitigation. The growing availability of geospatial and temporal data presents opportunities to develop AI and data science-based predictive models in the disaster domain. This work explores potentials and limitations in developing an AI-based LEWS using rainfall and soil moisture data from 2020 to 2023. We used 24-h rainfall and soil moisture data at 10 × 10 km resolution from the global precipitation measurement mission and National Snow and Ice Data Center (NSIDC), respectively. Machine learning models including support vector machine, random forest, gradient boosting (GB), XGBoost (XGB), and multi-layer perceptron were trained for grid-level landslide prediction. XGB consistently outperformed others, balancing precision and recall during the January–July 2024 test period. It achieved a precision of 0.8518, recall of 0.8041, F1-score of 0.8272 for the landslide class, and an AUC of 0.8973—demonstrating its effectiveness in minimizing false negatives, which is vital for early warning. Despite daily performance fluctuations, XGB showed reliable average results. However, lower recall compared to the majority class indicates ongoing challenges due to class imbalance, data gaps, and resolution limits. Finally, deployment of an open-source geo-portal for visualizing and monitoring predictions underscores the practical relevance of this work.
Chutia et al. (Wed,) studied this question.