Artificial Intelligence Techniques for Landslides Prediction Using Satellite Imagery

Landslides are among the most devastating natural hazards, causing significant damage to infrastructure, environment, and human lives. Accurate and timely detection of landslide‑prone areas is critical for hazard mitigation, early warning, and disaster management. In this study, we investigate the efficacy of modern deep‑learning segmentation techniques for automated landslide detection from high‑resolution satellite imagery. We utilize the publicly available Bijie Landslide Dataset — a dataset curated for landslide mapping via remote sensing — as well as another annotated collection from Landslide4Sense, for model development and validation.

Leveraging two complementary architectures — a modified U-Net+ (an enhanced U‑Net variant) and YOLOv8-Seg — our pipeline performs pixel-level segmentation to delineate landslide scars and boundaries. The U‑Net+ model captures multi-scale spatial context and refines boundary detection, while YOLOv8‑Seg enables efficient detection and segmentation even in complex terrain with heterogenous textures. We apply standard data preprocessing and augmentation steps — including cropping, normalization, flipping, and rotation — to improve model robustness under varying seasonal, lighting, and topographic conditions.

Experimental results demonstrate that the proposed segmentation models yield high accuracy in landslide delineation. Quantitative evaluation using metrics such as precision, recall, F1‑score, and intersection-over-union (IoU) indicates that U‑Net+ attains strong segmentation performance, while YOLOv8‑Seg shows competitive boundary detection with good generalization across different regions. Comparison with baseline and existing methods suggests that our approach improves both the delineation quality and computational efficiency.

This work highlights the potential of AI‑backed segmentation methods in remote sensing for landslide hazard mapping and risk assessment. The results can support early detection, monitoring, and mitigation strategies — especially in regions prone to landslide disasters. Future work will explore integrating additional spatial data (e.g., DEM, slope, vegetation) and temporal analysis to extend the approach toward predictive landslide susceptibility modelling.

Keywords: Landslide prediction, satellite imagery, deep learning, U-Net+, YOLOv8-Seg, segmentation, remote sensing, hazard mapping, AI techniques, disaster management, spatial data, landslide detection, machine learning, risk assessment, terrain analysis, model evaluation, precision, recall, F1-score, intersection-over-union (IoU).