Superpixel Fused Pixel Context Network for Precise Surface Water Mapping Using Multispectral Satellite Images

Precise delineation of surface water is essential for monitoring its status. Traditional indices-based methods suffer from the problem of falsely positive detection with a similar appearance, and finding a threshold value is also a challenging task. Deep learning (DL)-based methods need to handle high-class balance and pixel neighbor details for accurate surface water mapping. In this work, we proposed a novel DL-based Superpixel Fused Pixel context Network (SFPNet) with a lightweight Pixel Context Block and a Superpixel Block to extract surface water from satellite images. An exhaustive ablation study is presented to validate the effectiveness of our work. SFPNet has produced better results compared to the existing traditional and deep learning-based approaches.