Hyperspectral change detection (HCD) aims to recognize altered areas between hyperspectral images (HSIs) captured at different times, which is one of the crucial research areas in remote sensing. In recent years, convolutional neural networks (CNNs) and transformer-based models have been popularly exploited for HCD. However, these models are based on the fixed-weight linear transformation, which struggles to effectively model the intricate spectral-spatial relationships inherent in HSIs. Meanwhile, these methods usually neglect the feature distribution discrepancy induced by different environmental factors. To alleviate these issues, this article proposes a novel spectral-spatial-temporal Kolmogorov-Arnold network (SSTKAN) for HCD. First, a spectral-spatial Kolmogorov-Arnold network is proposed to extract the spectral-spatial features from the input bitemporal HSIs. Then, a 3-D Kolmogorov-Arnold network (3-D KAN) is exploited to extract the difference and temporal features. Next, a second-order statistical alignment method is proposed to reduce the feature distribution discrepancy of the same ground objects from bitemporal features. Finally, a multiscale feature enhancement (MSFE) module is designed to strengthen the feature representation by aggregating information from multiple receptive fields, followed by a fully connected layer to generate the final detection map. Experiments on four public hyperspectral datasets demonstrate that the proposed SSTKAN outperforms other advanced change detection (CD) approaches in both qualitative and quantitative results. The code of this article is available at https://github.com/PuhongDuan/SSTKAN.
Duan et al. (Thu,) studied this question.