Frequency-Domain Decoupling Method for Hyperspectral Anomaly Detection

In recent years, deep learning techniques support hyperspectral anomaly detection (HAD). However, existing methods based on background reconstruction and anomaly suppression overlook a fundamental characteristic: backgrounds contain low-frequency features, while anomalies manifest as highfrequency components. To address this issue, this study proposes a hyperspectral anomaly detection algorithm based on Fourier frequency-domain decoupling. First, an adaptive Fourier decomposition module is constructed to separate high-frequency and low-frequency components while preserving target low-frequency information; low-frequency components reconstruct the background, and high-frequency components are potential anomaly regions. Second, an improved Low-Frequency Adaptive Fusion Transformer (LFAT) is introduced into the low-frequency reconstruction branch, ensuring accurate background reconstruction through adaptive fusion of encoded features and reconstruction branch features. Finally, anomaly detection is accomplished using the proposed comprehensive Mahalanobis distance detector.