Frequency-Aware Kernel-Centered Prior Attention Network for Hyperspectral Image Classification

Abstract

Hyperspectral image (HSI) classification remains challenging because local discriminative cues around land-cover transitions, mixed pixels, and abrupt material boundaries are often under-modeled, while global attention can be distracted by spatially disconnected yet spectrally similar regions. To address these issues, this article proposes a frequency-aware kernel-centered prior attention network (FKCPA-Net) for HSI classification. The network first builds multiscale spectral–spatial representations to capture spectral continuity at different granularities. It then enhances local discrimination through a frequency-aware spectral–spatial fusion strategy that combines 3-D discrete cosine transform (DCT) high-pass enhancement, channel recalibration, and cross-group spectral–spatial interaction, thereby strengthening boundary, texture, and mixed-pixel representations. To improve global context modeling, a kernel-centered prior attention (KCPA) mechanism introduces complementary structural priors and a relative spatial-distance bias into dual-path attention and selectively preserves strongly correlated token interactions during normalization. Experiments on the Indian Pines (IPs), Botswana (BS), WHU-Hi-LongKou (LK), and HuangHeKou (HHK) datasets yield overall accuracies of 99.195%, 99.846%, 98.992%, and 98.111%, respectively. Additional analyses show that FKCPA-Net produces more spatially coherent classification maps and remains robust under limited training samples, demonstrating its effectiveness for complex HSI scenes. The code is available at https://github.com/DEVM369/FKCPA-Net