Flood Mapping Using Contrastive Learning and Semantic Segmentation

Floods are among the most destructive natural disasters and require accurate mapping for emergency response and mitigation. However, the lack of labeled data in remote sensing makes supervised approaches challenging. This paper proposes a semi-supervised flood segmentation model that leverages contrastive pretraining on synthetic aperture radar imagery. The approach transfers learned representations into a segmentation model trained with limited labeled data. We present a proof of concept to validate the performance of both the contrastive and segmentation models. The experiments in four regions (Nebraska, Bangladesh, Florence, and North Alabama) show promising results, achieving an average intersection over union of 0.52. The results suggest that contrastive learning can enhance flood detection in data scarce environments.