Consistency Regularization for Semi-Supervised Semantic Segmentation of Flood Regions from SAR Images

As one of the most powerful natural catastrophes, floods pose serious risks to people's lives, the integrity of infrastructure, and agricultural landscapes, which increases the toll they take on the economy and society. As a result, it becomes essential to continuously monitor these areas of vulnerability in order to support effective disaster response and mitigation efforts. Accurately defining the extent of floods is a problem for traditional flood mapping approaches, which emphasizes the vital need for modern technologies such as Synthetic Aperture Radar (SAR) imaging. Additionally, there is a need to develop computer-aided tools specifically designed for automatically identifying areas that are vulnerable to flooding using SAR data. Nonetheless, the lack of consistent large datasets presents a barrier that prevents these algorithms from progressing and being used in real-world scenarios. For this reason, the present study aims to develop a semi-supervised semantic segmentation algorithm for accurate flood region delineation in SAR data. In particular, the paper proposes labeling unannotated instances of data using a pseudo-label generation strategy. In order to accomplish this, the study suggests using a self-supervised trained teacher model to generate pseudo-labels and speed up the training procedure. The teacher model is then trained with a student model to efficiently extract features from the labeled data. Furthermore, the study presents a new semantic segmentation technique that uses convolutional neural networks to automatically identify flooded areas in SAR images. A comprehensive assessment conducted on publicly available datasets produces promising results. These results confirm the usefulness and possible relevance of the suggested methodology in enhancing efforts related to flood zone identification and management.