A Multi-Modal Parallel Deep Learning Framework Integrating Attention U-Net and Bidirectional LSTM for Spatio-Temporal Landslide Prediction

Abstract

Landslides are among the most significant natural hazards in mountainous regions, particularly under intense rainfall conditions. Accurate prediction requires the integration of both static terrain characteristics and dynamic hydrometeorological factors. However, many existing approaches rely primarily on spatial variables or aggregated rainfall indicators, limiting their ability to represent spatio-temporal relationships associated with slope failure. This study proposes a dual-stream multi-modal deep learning framework for spatio-temporal land-slide prediction in the Kerala Western Ghats, India. The framework combines an Attention U-Net branch for spatial feature extraction and a Bidirectional Long Short-Term Memory (BiLSTM) branch for temporal sequence modeling. The spatial input consists of a seven-channel tensor derived from Sentinel-2 spectral bands, the Normalized Difference Vegetation Index (NDVI), elevation, and slope layers, while the temporal input comprises a 30-day antecedent rainfall sequence obtained from the Climate Hazards Group InfraRed Precipitation with Stations (CHIRPS) dataset. Feature representations extracted from the spatial and temporal branches are integrated through a late-fusion mechanism and subsequently used for binary classification. Model performance was evaluated using an independent geographic holdout dataset. The proposed framework achieved an accuracy of 94.79%, precision of 89.74%, recall of 97.22%, F1-score of 93.33%, and an AUC-ROC value of 0.9757. Comparative experiments with Random Forest, ResNet-18, and unidirectional LSTM models demonstrated improved predictive performance of the proposed architecture. The results indicate that the integration of spatial and temporal information can enhance landslide prediction in data-driven frameworks for monsoon-dominated mountainous environments.