DCT-Mamba3D: Spectral Decorrelation and Spatial-Spectral Feature Extraction for Hyperspectral Image Classification

TL;DR

This work tackles spectral redundancy and complex spatial–spectral dependencies in hyperspectral image (HSI) classification by introducing DCT-Mamba3D, a three-module framework that combines a 3D Spatial-Spectral Decorrelation Module (3D-SSDM) with 3D DCT basis functions, a bidirectional 3D-Mamba module for spatial–spectral interaction, and a Global Residual Enhancement (GRE) module for stable learning. The 3D-SSDM decorrelates spatial and spectral information, producing a frequency-domain representation $X_{ ext{freq}}$; the 3D-Mamba module then models interactions across these decorrelated components, and GRE stabilizes representations via $F_{ ext{out}} = y_{ ext{mamba}} + \alpha X_{ ext{freq}}$, optimized by $\\mathcal{L} = \mathcal{L}_{\text{CE}} + \lambda \mathcal{L}_{\text{reg}}$. Empirical results on Indian Pines, KSC, and Houston2013 show state-of-the-art performance, with notable gains in challenging scenarios such as same object with different spectra and different objects with similar spectra, along with faster convergence and robustness under limited data. The proposed approach demonstrates that frequency-domain decorrelation combined with a bidirectional state-space model yields improved feature separability and classification accuracy, offering practical benefits for remote sensing tasks where spectral redundancy is prevalent.

Abstract

Hyperspectral image classification presents challenges due to spectral redundancy and complex spatial-spectral dependencies. This paper proposes a novel framework, DCT-Mamba3D, for hyperspectral image classification. DCT-Mamba3D incorporates: (1) a 3D spectral-spatial decorrelation module that applies 3D discrete cosine transform basis functions to reduce both spectral and spatial redundancy, enhancing feature clarity across dimensions; (2) a 3D-Mamba module that leverages a bidirectional state-space model to capture intricate spatial-spectral dependencies; and (3) a global residual enhancement module that stabilizes feature representation, improving robustness and convergence. Extensive experiments on benchmark datasets show that our DCT-Mamba3D outperforms the state-of-the-art methods in challenging scenarios such as the same object in different spectra and different objects in the same spectra.

Figures (6)

• Figure 1: Spectral response functions illustrating HSI classification challenges due to spectral redundancy and high correlation.
• Figure 2: DCT-Mamba3D framework.
• Figure 3: 3D Spatial-Spectral Decorrelation Module (3D-SSDM), applying 3D DCT basis functions to decompose the image into independent frequency components, aiding decorrelation and feature extraction.
• Figure 4: Spearman correlation heatmaps on the Indian Pines dataset, comparing (a) 2D-CNN, (b) HiT (Transformer-based), and (c) DCT-Mamba3D.
• Figure 5: t-SNE visualization of feature embeddings across models on the KSC dataset.