CrossKAN: A bivariate cross KAN model for hyperspectral change detection

Hyperspectral change detection (HCD) is a critical remote sensing approach for monitoring land surface changes. Despite notable progress, state-of-the-art HCD methodologies encounter difficulties in modeling the high-dimensional, nonlinear spectral characteristics of hyperspectral imagery when comparing pre- and post-change imagery. To address these limitations, we propose a novel deep learning architecture, termed Cross Kolmogorov–Arnold Network (CrossKAN), for accurate and interpretable HCD. The CrossKAN model is predicated on the functional decomposition theory of Kolmogorov–Arnold Networks, a theoretical framework that enables compact and mathematically grounded modeling of complex spectral relationships. A Siamese architecture is employed to process bi-temporal image patches, enabling robust feature extraction by KAN layers based on Chebyshev polynomials. Next, deep features are fused in the CrossKAN layer, and are fed to the subsequent KAN layers to discriminate between change and no-change locations. CrossKAN's performance was assessed using four benchmark datasets in different geographical locations with divergent context and change classes. CrossKAN outperformed state-of-the-art HCD models, including SSTFormer, DBS 3 TAN, ML-EDAN, and MSDFFN, and achieved an overall accuracy of >94%. Low missed detection and false alarm rates demonstrate CrossKAN's superior effectiveness and generalization in complex regions. • Novel CrossKAN architecture integrates Kolmogorov-Arnold Networks with Siamese framework for mathematically principled hyperspectral change detection modeling. • Chebyshev polynomial-based KAN layers enable explicit spectral response function modeling while preserving computational efficiency and interpretability. • Achieves >94% accuracy across four benchmark datasets, significantly outperforming state-of-the-art methods including SSTFormer and DBS 3 TAN.