Improved lightweight water body recognition model using Mobile-Former network

ABSTRACT To address high computational costs and prolonged training in conventional water body recognition models (e.g., Swin Transformer, DeepLabV3+), we propose a lightweight network replacing DeepLabV3 + 's backbone with Mobile-Former. Mobile-Former achieves bidirectional fusion of local and global features via parallel MobileNetV3 and Transformer blocks. The model extracts a shallow feature and two deep features, with one deep feature processed by dilated convolution before fusion with the shallow feature and the remaining deep feature. This design reduces FLOPs from 219,300.75 to 3,727.14 M and parameters from 351.56 to 12.21 M. Experimental results show average IoU scores of 67.98, 69.86, and 91.81% on three datasets. The lightweight architecture enables real-time processing and suits resource-constrained environments. Validation on 2020 Chaohu flood data demonstrates 98.79% accuracy in flood inundation extraction, highlighting its practical efficacy for disaster response applications.