Learnable Frequency-Domain Phase Residual and Graph-Contrastive Dynamic Convolution Network for Hyperspectral Anomaly Detection

Hyperspectral anomaly detection aims to identify sparse pixels with significant spectral or spatial differences from the surrounding background without prior target spectral information, and it has significant application value in key fields such as military reconnaissance and environmental monitoring. Existing methods suffer from insufficient frequency-domain feature extraction, static multiscale spatial modeling, and inadequate spectral–spatial feature interaction mechanisms. These limitations hinder their ability to adapt to complex scenarios involving target spectral variations, diverse target scales, and intricate spatial structures. Therefore, we propose a learnable frequency domain phase residual and graph-contrastive dynamic convolution (LFGDC) network for hyperspectral anomaly detection that integrates learnable frequency-domain modeling, multiscale dynamic convolution, and graph-structured contrastive constraints. First, a learnable frequency domain phase residual block (LFPRB) is introduced to explicitly exploit spectral frequency-domain characteristics. Second, a multiscale dynamic convolution module (MDCM) generates input-dependent convolution kernels to effectively adapt to spatial representations of targets at different scales. Finally, a graph-contrastive structure fusion module (GCSFM) is constructed to uncover global-local topological correlations and enhance feature discriminability. This is combined with adaptive frequency-domain modeling and multiscale feature fusion strategies to bolster model robustness. Extensive experimental results demonstrate that this method achieves superior detection performance compared to baseline methods on both public and produced datasets, providing an effective solution for hyperspectral anomaly detection in complex scenarios.