FFTNet2D: A Novel Frequency Domain Framework for Hyperspectral Image Classification

Abstract

Hyperspectral image (HSI) classification remains challenging because accurate recognition requires both global contextual modeling and computationally efficient processing of high-dimensional spectral–spatial data. Transformer-based methods can capture long-range interactions, but their quadratic token-mixing cost limits scalability; convolutional networks are efficient but remain constrained by local receptive fields. To address this tradeoff, we propose FFTNet2D, a frequency-domain architecture that replaces self-attention with fast Fourier transform (FFT) for efficient global feature modeling. The framework introduces a frequency domain channel attention (FDCA) module that modulates the amplitude spectrum of learned hyperspectral feature maps, enhancing discriminative frequency responses while suppressing redundant and noisy components. A dual-path design further integrates local spatial representations from depthwise convolutions with global frequency representations extracted by the FFT. Experiments on the Indian Pines and WHU-Hi-LongKou datasets show that FFTNet2D achieves favorable accuracy–efficiency tradeoffs, with particularly strong performance on the more challenging UAV-borne WHU-Hi-LongKou agricultural scene.