SSLANet: A Frequency–Spatial Visual System for Efficient Hyperspectral Image Classification

Abstract

Hyperspectral image classification requires a compact representation that can jointly capture discriminative spectral cues, local spatial structures, and efficient inference behavior under limited annotations. Existing CNN-based methods provide effective local pattern extraction but are less capable of modeling long-range spectral dependencies. Transformer-based models improve global interaction at the cost of considerable computational overhead, whereas recent state-space models usually formulate hyperspectral data as spatial–spectral sequences and still leave transform-domain spectral variations insufficiently explored. To address these limitations, we propose a spectral–spatial learning adaptive network (SSLANet), an efficient frequency–spatial learning network for hyperspectral image classification. The core component is a residual frequency–spatial layer that integrates an adaptive spectral block (ASB) and an interactive convolution block (ICB). ASB performs learnable complex-valued spectral modulation together with input-dependent frequency-energy selection to enhance material-sensitive spectral responses, while ICB propagates the filtered features into local neighborhoods through multiplicative spatial–channel interaction. This design differs from fixed transform-domain preprocessing in that the frequency response is conditioned on the current input feature distribution through a soft mask and an effective transfer function in the latent frequency domain. Extensive experiments on three public hyperspectral benchmarks demonstrate that SSLANet achieves competitive classification accuracy with favorable computational efficiency. The source code is available at https://github.com/NELUJA86/SSLANet