Frequency-Aware Bias Calibration Attention Network for Hyperspectral Image Classification

Abstract

Recently, self-attention mechanisms have significantly advanced hyperspectral image classification (HSIC) by enabling effective modeling of long-range spatial–spectral dependencies. However, existing methods often suffer from blurred land-cover boundaries and insufficient discriminability between spectrally similar classes due to redundant global interactions and inadequate preservation of fine-grained spatial details. To address these challenges, we propose a novel frequency-aware bias calibration attention network (FBCANet) for HSIC. Specifically, we design a frequency-domain spatial–spectral global aggregation module to model global contextual dependencies in the frequency domain. Within this module, a global frequency enhancement filter (GFEF) explicitly decouples complex frequency representations to adaptively suppress redundant interference caused by similar spectral signatures across different categories. Furthermore, we introduce a bias-calibrated global attention (BCGA) module, which compensates for the representation bias introduced by global attention modeling. This module consists of a global feature aggregation backbone and an explicit bias-calibration branch that extracts fine-grained spatial–spectral cues and injects them into the global interaction process, thereby enhancing interclass discriminability. In addition, a dual-path frequency-aware interaction module is proposed to capture high-frequency information and dynamically compensate for the loss of detailed textures and complex land-cover boundaries. Extensive experiments on three benchmark hyperspectral datasets demonstrate that the proposed method learns highly discriminative feature representations and consistently outperforms state-of-the-art HSIC methods. The source code for this work will be publicly available at https://github.com/szq0816/FBCANet