Unsupervised deep learning for environmental risk monitoring: Landslide detection from multi-resolution remote sensing imagery

This study assesses three self-supervised learning (SSL) models including SimCLR, SwAV, and DINOv2 for landslide segmentation using Sentinel-2 and Landsat 8 imagery. The evaluation leverages the Landslide4Sense dataset and a custom Canadian dataset encompassing Yukon, the Northwest Territories, British Columbia, and Northern Quebec. Embedding analysis showed SSL models effectively distinguished landslide features, with DINOv2 yielding high similarity (0.56–0.842) for landslide images and low/negative scores (<0.015, -0.175 to -0.181) for dissimilar land covers/noise. Pretrained on unlabeled multispectral data and fine-tuned with 1% and 10% labeled data, DINOv2 outperformed SimCLR, SwAV, and a supervised U-Net baseline, achieving F1-scores of 0.87 (1% data) and 0.94 (10% data). SimCLR and SwAV scored 0.77 and 0.83 (1% data), improving to 0.83 and ∼0.88–0.90 (10% data), while supervised U-Net reached 0.84. In Canadian regions, DINOv2 excelled with F1-scores of 0.72–0.91 across diverse landslide types, followed by SwAV (0.64–0.90), with Sentinel-2 generally outperforming Landsat 8, except for permafrost landslides where Landsat 8 achieved 0.79 vs. 0.72. Compared to prior studies, DINOv2 surpassed supervised baseline and other SSL models, driven by its transformer-based architecture and strategic band selection. Despite limitations with 128×128 patches and dataset imbalances, SSL models prioritized high recall, ensuring robust detection. These results enable near-real-time landslide mapping in data-scarce regions using freely available Sentinel-2/Landsat imagery, reducing dependency on expensive manual labeling, and supporting rapid post-event assessment, early warning integration, and resource allocation in disaster response workflows.