Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

WaveMamba: Spatial-Spectral Wavelet Mamba for Hyperspectral Image Classification

Muhammad Ahmad, Muhammad Usama, Manuel Mazzara, Salvatore Distefano

TL;DR

Hyperspectral image classification often faces data-efficiency and compute constraints. The paper proposes WaveMamba, which fuses Haar wavelet-based multi-resolution spatial-spectral features with the Spatial-Spectral Mamba backbone and a State-Space Model to capture local textures, global context, and temporal dependencies. Key ideas include dividing HSI into overlapping 3D patches, generating spectral and spatial tokens, applying a gate-driven refinement, and decomposing features into Haar wavelet subbands for state-space processing, with L2 regularization at the classifier. Empirical results on University of Houston and Pavia University show notable accuracy gains (4.5% and 2.0% respectively) over strong baselines, with WaveMamba outperforming CNN-, Transformer-, and Mamba-based methods in OA, AA, and kappa, especially under limited training data.

Abstract

Hyperspectral Imaging (HSI) has proven to be a powerful tool for capturing detailed spectral and spatial information across diverse applications. Despite the advancements in Deep Learning (DL) and Transformer architectures for HSI classification, challenges such as computational efficiency and the need for extensive labeled data persist. This paper introduces WaveMamba, a novel approach that integrates wavelet transformation with the spatial-spectral Mamba architecture to enhance HSI classification. WaveMamba captures both local texture patterns and global contextual relationships in an end-to-end trainable model. The Wavelet-based enhanced features are then processed through the state-space architecture to model spatial-spectral relationships and temporal dependencies. The experimental results indicate that WaveMamba surpasses existing models, achieving an accuracy improvement of 4.5\% on the University of Houston dataset and a 2.0\% increase on the Pavia University dataset.

WaveMamba: Spatial-Spectral Wavelet Mamba for Hyperspectral Image Classification

TL;DR

Hyperspectral image classification often faces data-efficiency and compute constraints. The paper proposes WaveMamba, which fuses Haar wavelet-based multi-resolution spatial-spectral features with the Spatial-Spectral Mamba backbone and a State-Space Model to capture local textures, global context, and temporal dependencies. Key ideas include dividing HSI into overlapping 3D patches, generating spectral and spatial tokens, applying a gate-driven refinement, and decomposing features into Haar wavelet subbands for state-space processing, with L2 regularization at the classifier. Empirical results on University of Houston and Pavia University show notable accuracy gains (4.5% and 2.0% respectively) over strong baselines, with WaveMamba outperforming CNN-, Transformer-, and Mamba-based methods in OA, AA, and kappa, especially under limited training data.

Abstract

Hyperspectral Imaging (HSI) has proven to be a powerful tool for capturing detailed spectral and spatial information across diverse applications. Despite the advancements in Deep Learning (DL) and Transformer architectures for HSI classification, challenges such as computational efficiency and the need for extensive labeled data persist. This paper introduces WaveMamba, a novel approach that integrates wavelet transformation with the spatial-spectral Mamba architecture to enhance HSI classification. WaveMamba captures both local texture patterns and global contextual relationships in an end-to-end trainable model. The Wavelet-based enhanced features are then processed through the state-space architecture to model spatial-spectral relationships and temporal dependencies. The experimental results indicate that WaveMamba surpasses existing models, achieving an accuracy improvement of 4.5\% on the University of Houston dataset and a 2.0\% increase on the Pavia University dataset.
Paper Structure (8 sections, 6 equations, 4 figures, 4 tables)

This paper contains 8 sections, 6 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Proposed Methodology
  3. Experimental Results and Discussion
  4. Patch Size and Train-Validation-Test Samples
  5. Impact of Wavelet Transformation
  6. Experimental Results with CNN-based Models
  7. Comparative Results with Transformer and Mamba
  8. Conclusions

Figures (4)

  • Figure 1: The HSI cube is divided into overlapping 3D patches further split into spectral and spatial tokens, enhanced using a gate mechanism, and transformed with Haar wavelets into four subbands that capture different frequency components and spatial features. The subbands are concatenated into a new 3D representation, which the Mamba architecture uses to capture spatial-spectral relationships and temporal dependencies.
  • Figure 2: Accuracy and loss trends of CNN-based methods.
  • Figure 3: The proposed WaveMamba achieves OA=97.30% showing competitive performance. All these results are compiled using $10 \times 10$ patch size with 25% training samples for all competing methods.
  • Figure 4: The proposed WaveMamba achieves OA=96.58% showing competitive performance. All these results are compiled using $10 \times 10$ patch size with 25% training samples for all competing methods.