AKAN-FSL: Adversarial Kolmogorov–Arnold Network for Cross-Domain Few-Shot Hyperspectral Image Change Detection

Abstract

Few-shot learning (FSL) provides a promising solution for reducing annotation cost in hyperspectral image change detection (HSI-CD). However, most existing FSL approaches focus on single-modality or homogeneous domains, limiting their ability to handle domain shift. Moreover, many feature extractors rely on fixed kernels or fixed attention operators, making them less effective at modeling the nonlinear spatial-spectral relationships under distribution mismatch. To address these challenges, this paper proposes an adversarial Kolmogorov-Arnold network for cross-domain few-shot HSI-CD (AKAN-FSL). The method alternates episodic training between source and target domains to transfer prior knowledge while employing adversarial training to mitigate domain shift. First, very-high-resolution imagery (VHRI) is used as the source domain, while HSI datasets serve as the target domains. Then, a KAN feature generation module (KAN-FGM) is designed as a shared feature generator that simultaneously supports FSL and adversarial domain alignment. To complement KAN-FGM, a multi-scale discriminator (MSD) is designed to fuse global and local features. Through adversarial alignment between KAN-FGM and MSD, the distribution discrepancy between heterogeneous domains is reduced. Experiments on three HSI datasets demonstrate that AKAN-FSL achieves competitive performance under a few labeled samples compared with other FSL methods.