SwinSegFormer: Advancing Aerial Image Semantic Segmentation for Flood Detection

Abstract

Semantic segmentation of aerial images is essential for unmanned aerial vehicle (UAV) applications in disaster management, particularly for identifying the flood-affected areas. Traditional techniques face challenges in capturing global semantic information due to their limited receptive fields, and high computational requirement. To address these issues, we propose a novel transformer-based model named SwinSegFormer, which feature a hierarchical encoder that efficiently generates multi-scale high-resolution features along with a lightweight decoder to reduce computational overhead. The proposed model is trained on FloodNet dataset and demonstrates efficient performance on challenging classes such as vehicles, pools, and flooded and non-flooded roads, which are crucial for effective disaster management. Additionally, we developed a post-processing module to categorize areas into flooded and non-flooded. The model achieves a validation mIoU of 75.1%, mDice of 85.4%, and mACC of 87.1%, representing a 10-12% improvement over state-of-the-art vision transformer-based methods. The effectiveness of model is further evaluated on real-world unlabeled flood imagery, highlighting its potential for supporting first aid activities during floods.