The intelligent and efficient analysis of agricultural surface water is essential for agricultural irrigation, flood prevention, and resource utilization. However, accurate extraction from high-resolution optical remote sensing images is challenging due to agricultural heterogeneous terrains, dynamic water body scales, and diverse surrounding vegetation. To address these challenges, this paper introduces a novel architecture called the Complex Rural Water Extraction Network (CRWENet). The CRWENet is designed to enhance the understanding of contextual information and suppress random background noise, which comprises four key components including feature encoder, feature decoder, context information enhancement module, and foreground detail attention module. The context information enhancement module harnesses the long-range dependency-capturing capabilities of Transformer, merging multi-scale features to enhance global information perception. Meanwhile, the foreground detail attention module facilitates adaptive suppression of intricate background noise and accurate delineation of agricultural surface water boundaries by harmonizing interactions across channel, height, and width dimensions. Extensive experimentation on the LoveDA dataset and the GLH-water dataset validates the effectiveness of CRWENet. The results demonstrate that the proposed CRWENet outperforms the existing classical semantic segmentation networks and the mainstream water body extraction methods, achieving Intersection over Union (IoU) scores of 70.93% and 80.55% respectively.
Read full abstract