A Deep Learning Framework for Landslide Detection in Support of Early Warning Systems

Abstract

Landslides are sudden and destructive hazards that have a significant impact on human, economic, and environmental aspects, and their complex triggering factors make them particularly difficult, underscoring the importance of reliable Landslide Early Warning Systems (LEWS). Traditional approaches are often based on empirical thresholds or statistical models. They provide valuable insights but are generally constrained in accuracy and transferability across different regions. Recent advances in deep learning offer new opportunities by enabling the automatic extraction of spatial and geomorphic features from high-resolution imagery to digital elevation data. This study proposes a framework that incorporates different deep learning methods and a landslide dataset to enhance landslide detection in support of LEWS. The model integrates spectral and terrain variables, applies data augmentation to mitigate class imbalance, and is evaluated with precision, recall, F1-score, and IoU against conventional classifiers. Results demonstrate that the proposed approach improves detection accuracy and robustness in heterogeneous terrain, indicating its potential as a scalable and transferable component of operational early warning systems and risk management.