Comparative Study of Water Body Segmentation Models Based on SegFormer and K-Net

To address the widespread issues of boundary blurring and low accuracy in identifying small water bodies during water body recognition, state-of-the-art semantic segmentation models, SegFormer and K-Net, were utilized to construct water body recognition models for comparative analysis. The experimental results indicate that when water body areas exhibit high continuity and distinct boundary features, the recognition performances of both models are fundamentally identical. However, the K-Net model demonstrates stronger robustness than the SegFormer model in the segmentation of slender water bodies. An analysis of evaluation metrics reveals that the Intersection over Union (IoU), Accuracy (Acc), Dice coefficient, and Kappa coefficient of the K-Net model reach 97.14%, 99.13%, 98.55%, and 97.93%, respectively. All these metrics surpass those of the SegFormer model, indicating that the K-Net water body recognition model possesses superior generalization ability and robustness. Consequently, it is more suitable for addressing the identification challenges of complex and small water bodies.