A Lightweight Network With Embedded Soft Constraints on Approximate Spectral Features for Real-Time Water Body Segmentation in Remote Sensing Images

Real-time extraction of water bodies from remote sensing images captured by unmanned aerial vehicles (UAVs) or satellites is challenging due to the difficulty in performing precise atmospheric correction and other preprocessing steps. In this article, we propose a lightweight network with embedded soft constraints on approximate spectral features for real-time water body segmentation in remote sensing images. First, we introduce approximate spectral feature indices as auxiliary interband feature information. A lightweight pseudo-siamese feature extraction network (LPSE) is designed to separately extract features from visible bands and the approximate spectral indices. Second, we develop an approximate spectral feature soft constraint fusion mechanism (ASFC) that utilizes spatial attention to selectively fuse effective target features from the visible bands and approximate spectral indices. Third, we incorporate an atrous spatial pyramid pooling module for edge feature enhancement within a self-distillation edge-aware lightweight decoder. Finally, the proposed network is accelerated and quantized using TensorRT and deployed on the embedded device Jetson Orin NX. Experimental results show that the model achieves an intersection over union accuracy of 70.74% on the FloodNet dataset and 92.26% on the GF-FloodNet dataset. With only 0.22 million parameters and a computational cost of 0.32 GFLOPs, the inference time per image is 6.45 ms. The proposed method demonstrates significant advantages in segmentation accuracy and computational efficiency, making it highly promising for real-time water body segmentation on edge computing platforms, such as UAVs or satellites.