Enhancing Landslide Detection with U-Net and Focal Loss Benchmarking Single Models, Ensembles, and Knowledge Distillation on Landslide4sense Dataset
Year: 2026
Type: conference-paper
Abstract: Landslide detection is critical for disaster risk reduction, particularly in landslide-prone regions of Asia where approximately 75% of fatal landslide events occur. This study presents a comprehensive benchmarking of deep learning architectures for landslide segmentation using the Landslide4Sense dataset. We evaluate five approaches: CNN, U-Net, Attention U-Net, ResNet-style network, and knowledge distillation from multiple teachers. The Attention U-Net model achieves the best overall performance with 70.90% precision, 99.57% recall, and 82.82% F1-score, representing a significant improvement over the baseline U-Net. The knowledge-distilled student model achieves 86.01% recall with only 66.66% precision, making it suitable for resource-constrained deployment. Our results demonstrate that Attention U-Net with Focal Loss provides the optimal balance for practical landslide detection systems, detecting 99.6% of all landslide pixels. The comparative analysis shows that no single architecture dominates all metrics; rather, the choice depends on application requirements. This study contributes a systematic evaluation of state-of-the-art semantic segmentation networks for landslide mapping, providing practitioners with clear guidance for model selection based on precision-recall trade-offs.
Authors: Jorton A. Tagud, Ramil Lumauag
Language: English
Subfield: Management, Monitoring, Policy and Law
Field: Environmental Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204147944