An improved U-shape Kolmogorov-Arnold network for bridge detection of remote sensing images

Accurate bridge detection in remote sensing images is crucial for water resource management, especially in flood-prone areas. However, traditional deep learning models such as U-Net struggle to effectively identify bridges due to their diverse morphologies, small size relative to the overall image, and the high-resolution nature of remote sensing imagery. This study presents an improved U-KAN model, which integrates the Kolmogorov-Arnold Network (KAN) with an enhanced U-Net framework and introduces a channel attention mechanism using the Squeeze-and-Excitation (SE) module. The proposed model addresses the challenges of class imbalance, boundary ambiguity, and multi-scale variability by incorporating KANLinear layers and replacing ReLU activations with LeakyReLU. Extensive experiments on a bridge dataset demonstrate that the proposed algorithm outperforms state-of-the-art models, achieving a higher Intersection over Union (IoU), precision, and recall. Comparative and ablation studies further validate the effectiveness of the proposed architectural modifications. This improved U-KAN provides a robust and efficient solution for automatic bridge extraction from remote sensing images, offering valuable support for water resource management applications.