Enhancing Landslide Prediction Through Advanced Transformer-Based Models: Integrating SAR Imagery and Environmental Data

Abstract

Landslides, a significant natural hazard driven predominantly by rainfall infiltration, pose a continuous threat to the terrain, buildings, and ecosystem of Caiazzo, located in southern Italy (Caserta). This research undertakes a novel approach to predict and understand the complex patterns of landslide deformations in this region. Our study leverages the advanced capabilities of COSMO-SkyMed satellite imagery, integrating a comprehensive dataset that includes factors such as elevation, slope, Topographic Wetness Index, Stream Power Index, geology, flow direction, curvature, Normalized Difference Vegetation Index, and several other relevant indices. We employ Transformer-based models, renowned for their effectiveness in capturing long-range dependencies within sequential data. The Transformer models in our study are adept at analyzing the temporal sequences of environmental factors and their intricate interactions, thereby offering a more nuanced understanding of the temporal patterns leading to landslides. These Transformer models are designed to process the entire sequence of satellite data obtained by the Coherent Pixels - Temporal Phase Coherence within the SUBSOFT package to ensure data integrity and precision, encompassing 132 images in ascending geometry and 143 in descending geometry, spanning from 2013 to 2021. By doing so, they efficiently identify critical patterns and dependencies over time, such as the interplay between rainfall events and subsequent soil and vegetation changes. The models are fine-tuned to our specific dataset, ensuring high precision and accuracy in landslide deformation prediction. Our evaluation metrics, including Mean Absolute Error, Root Mean Square Error, and R^2 score, demonstrate the superior performance of the Transformer-based approach, with significant improvements over the conventional Deep Learning model. The visual correlation of our predictions with actual landslide occurrences further corroborates the effectiveness of this method. This transformative approach not only enhances our understanding and predictive capability for landslides in Caiazzo but also sets a benchmark for landslide prediction in geologically vulnerable regions worldwide.