Ensemble teacher models for semi-supervised semantic segmentation of flood zones from synthetic aperture radar images

Floods cause significant economic devastation and result in the loss of human lives. Observation of flood-prone regions is essential for improved disaster management and mitigates the hazards to human life. Nevertheless, the process of manually monitoring these areas is laborious and demanding. In this context, Synthetic Aperture Radar (SAR) images can enhance the monitoring process and aid in disaster management. In addition, an automated system capable of analyzing these images has the potential to minimize human mistakes and eliminate subjective judgments. Deep learning models are commonly used in the literature for this purpose. However, the extent to which they may be used is restricted due to their reliance on extensive annotated datasets. Therefore, semi-supervised learning has become a widely discussed and researched issue in the academic world. Despite this, the advancement of the semi-supervised learning approach is difficult since it struggles with generalization and tends to provide inaccurate classifications. Given this information, the aim of this research study is to develop a semi-supervised learning method for accurately identifying flood areas using SAR images. The study proposes using ensemble teacher models to provide pseudo-labels for the unlabeled training data, which will then be used to train student models. In addition, a unique method for aggregating pseudo-labels is presented, which efficiently combines them and reduces noise in the predictions. The proposed model is evaluated both subjectively and statistically using a public dataset. Specifically, the method achieved a mean Intersection over Union (mIoU) of 0.63 and an F1-score of 0.82, reflecting a 3% improvement over comparative models. This highlights the efficacy of the proposed ensemble semi-supervised segmentation framework.