Ultralightweight progressive feature disentanglement and recomposition network for hyperspectral image classification

Deep Learning-based Hyperspectral Image Classification (DL-HSIC) demands highly discriminative spatial-spectral representation while maintaining a balance between computational practicality and classification performance. However, existing DL-HSIC approaches often suffer from excessive parameters and high inference latency, limiting deployment in resource-constrained scenarios. To address this challenge, a novel Ultralightweight Progressive Feature Disentanglement and Recomposition Network (ULite-FDRNet) is proposed in this study, which attains a superior trade-off among representation capability, parameter efficiency, and classification performance. ULite-FDRNet is built on a unified Feature Disentanglement and Recomposition (FDR) paradigm and comprises three compact modules: (1) FDRConv2D/3D is designed to efficiently extract local spatial details and preserve spectral dependencies. (2) Ultralight Axial Multiscale FDR 3D Convolution (ScaleFDR3D) introduces axial-sensitive kernels to enhance the multiscale diversity of spatial-spectral decoupling features. (3) TriAxial Synergistic FDR Attention Module (TriSFDR) disentangles attention into vertical, horizontal, and spectral three independent axes to acquire global spatial attention and coarse-fine-grained spectral attention. These components enable progressive feature learning from Shallow to Multiscale to Global. Extensive experiments on four HSI benchmarks demonstrate that ULite-FDRNet maintains superior accuracy while using substantially fewer parameters, enabling faster inference and practical deployment. Specifically, ULite-FDRNet achieves overall accuracies (OA) of 98.24%, 97.83%, 97.47%, and 99.57%, with merely 1.93 K, 2.05 K, 3.51 K, and 0.86 K parameters, respectively. The source code will be released at https://github.com/DelonKong/ULite_FDRNet.