Contrastive Self-Supervised Learning for Globally Distributed Landslide Detection
Citation
Ghorbanzadeh, O., Shahabi, H., Piralilou, S.T., Crivellari, A., Cué La Rosa, L.E., Atzberger, C., Li, J., Ghamisi, P. (2024). Contrastive Self-Supervised Learning for Globally Distributed Landslide Detection. IEEE Access, 12: 118453-118466. Link to paper
Abstract
The study addresses satellite imagery analysis challenges by leveraging self-supervised learning to handle unlabeled data. The researchers integrated SwAV within an auto-encoder framework using ResNet-18 architecture for landslide detection from multi-spectral imagery in the Landslide4sense benchmark dataset. Remarkably, their SSL model trained with only 1% labeled data matched the performance of ten state-of-the-art supervised methods using 100% labeled data. With 10% labeled data, it outperformed all ten fully supervised baselines, demonstrating significant potential for natural hazard mapping with limited labeled training samples.