DDFGFormer: Dual-Decoupled Frequency-Guided Transformer for Hyperspectral Image Classification

Abstract

Hyperspectral image (HSI) classification requires jointly capturing subtle spectral differences and complex spatial structures. Recent studies have introduced frequency-domain modeling to complement spatial–spectral representation learning. However, existing frequency-based methods still face some limitations. First, most methods mainly focus on amplitude or energy distributions while insufficiently exploiting phase information for structural localization and edge representation. In addition, raw phase information is sensitive to noise, smooth regions, and mixed pixels, which may introduce unreliable interference. Furthermore, spatial–frequency interaction is often implemented through simple posterior fusion, failing to use frequency-domain priors to actively guide spatial feature extraction. To address these limitations, we propose a dual-decoupled frequency-guided transformer, termed DDFGFormer, for HSI classification. In particular, we design a frequency-guided transformer block (FGTB) composed of dual-decoupled frequency attention (DDFA) and a frequency-enhanced feed-forward network (FEFFN). DDFA decouples the Fourier representation into amplitude and phase, and further decomposes the amplitude into high- and low-frequency components, thereby explicitly capturing amplitude- and phase-aware cues. To suppress unreliable phase responses, an amplitude-prior guided phase (APGP) module contained in DDFA is proposed to adaptively modulate phase representations and transform raw phase into stable structural priors. Furthermore, FEFFN injects frequency-domain information into feature propagation, enabling frequency priors to actively recalibrate spatial features rather than serving only as posterior fusion cues. Extensive experiments on four HSI datasets show that DDFGFormer outperforms nine state-of-the-art methods in overall accuracy (OA), average accuracy (AA), kappa, and boundary $F1$ -score (BF1), with OA gains of 1.65%, 1.63%, 1.25%, and 1.14% over the second-best method on Indian Pines (IPs), Pavia University (PU), Salinas (SAs), and WHU-Hi-LongKou (LK), respectively.