HydroKAN-Net: A Spectral-Gated Kolmogorov–Arnold Network With Boundary-Aware Cross-Scale Aggregation for Flood Segmentation in UAV Images

Abstract

Timely and accurate delineation of flooded regions from uncrewed aerial vehicle (UAV) imagery is critical for disaster response, damage assessment, and resource allocation. Yet flood segmentation remains difficult: water surfaces exhibit specular reflections, low-texture homogeneity, and diffuse, irregular boundaries against visually similar wet terrain, shadows, and debris. Convolutional U-Net variants are limited by local receptive fields, while transformer and state-space hybrids improve global context at the cost of heavy computation and weak fine-boundary recovery. To address these limitations, we propose HydroKAN-Net, a Kolmogorov–Arnold-based encoder–decoder that couples spline-driven nonlinear feature modeling with frequency- and boundary-aware refinement through three components. First, a spectral-gated KAN block (SG-KAB) performs a learnable frequency-domain decomposition before B-spline mapping, separating the low-frequency, homogeneous response of standing water from high-frequency boundary and clutter cues. Second, a boundary-aware cross-scale aggregation (BCSA) bridge replaces the additive skip connection, combining a Sobel-derived water-edge spatial gate with a channel gate driven by a shared multidepth global context vector to sharpen the land–water interface and suppress false background activations. Third, a water-aware frequency loss (WAFL), applied with multiscale deep supervision, augments Dice and cross-entropy terms with a spectral-consistency penalty that discourages fragmented masks. Across five independent runs, we evaluate HydroKAN-Net on four public UAV flood benchmarks, the flood semantic segmentation, FloodNet, Calgary, and Houston datasets, against U-Net, UNet++, TransUNet, Swin-UNet, Rolling-UNet, U-Mamba, and UKAN. The proposed model achieves the highest segmentation accuracy while remaining computationally efficient, and ablations confirm that SG-KAB, BCSA, and WAFL each contribute complementary gains.