Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DMS2F-HAD: A Dual-branch Mamba-based Spatial-Spectral Fusion Network for Hyperspectral Anomaly Detection

Aayushma Pant, Lakpa Tamang, Tsz-Kwan Lee, Sunil Aryal

TL;DR

DMS2F-HAD tackles hyperspectral anomaly detection by combining long-range spatial and spectral modeling in a lightweight, unsupervised dual-branch Mamba autoencoder. A learnable adaptive gated fusion module integrates spatial texture and spectral consistency per pixel, while a Spatial-Spectral Decoder enables reconstruction-based anomaly scoring via residuals. Across 14 benchmark datasets, the approach achieves a 14-dataset average AUC of 0.9878 and exhibits 4.6× faster inference with significantly fewer parameters than Transformer-based counterparts and other Mamba models. The method demonstrates strong generalization and efficiency, making it a practical solution for real-time HAD applications. The results highlight the effectiveness of using linear-time state-space models to replace attention-heavy mechanisms in hyperspectral processing, with the gate fused representation enhancing robustness across diverse scenes.

Abstract

Hyperspectral anomaly detection (HAD) aims to identify rare and irregular targets in high-dimensional hyperspectral images (HSIs), which are often noisy and unlabelled data. Existing deep learning methods either fail to capture long-range spectral dependencies (e.g., convolutional neural networks) or suffer from high computational cost (e.g., Transformers). To address these challenges, we propose DMS2F-HAD, a novel dual-branch Mamba-based model. Our architecture utilizes Mamba's linear-time modeling to efficiently learn distinct spatial and spectral features in specialized branches, which are then integrated by a dynamic gated fusion mechanism to enhance anomaly localization. Across fourteen benchmark HSI datasets, our proposed DMS2F-HAD not only achieves a state-of-the-art average AUC of 98.78%, but also demonstrates superior efficiency with an inference speed 4.6 times faster than comparable deep learning methods. The results highlight DMS2FHAD's strong generalization and scalability, positioning it as a strong candidate for practical HAD applications.

DMS2F-HAD: A Dual-branch Mamba-based Spatial-Spectral Fusion Network for Hyperspectral Anomaly Detection

TL;DR

DMS2F-HAD tackles hyperspectral anomaly detection by combining long-range spatial and spectral modeling in a lightweight, unsupervised dual-branch Mamba autoencoder. A learnable adaptive gated fusion module integrates spatial texture and spectral consistency per pixel, while a Spatial-Spectral Decoder enables reconstruction-based anomaly scoring via residuals. Across 14 benchmark datasets, the approach achieves a 14-dataset average AUC of 0.9878 and exhibits 4.6× faster inference with significantly fewer parameters than Transformer-based counterparts and other Mamba models. The method demonstrates strong generalization and efficiency, making it a practical solution for real-time HAD applications. The results highlight the effectiveness of using linear-time state-space models to replace attention-heavy mechanisms in hyperspectral processing, with the gate fused representation enhancing robustness across diverse scenes.

Abstract

Hyperspectral anomaly detection (HAD) aims to identify rare and irregular targets in high-dimensional hyperspectral images (HSIs), which are often noisy and unlabelled data. Existing deep learning methods either fail to capture long-range spectral dependencies (e.g., convolutional neural networks) or suffer from high computational cost (e.g., Transformers). To address these challenges, we propose DMS2F-HAD, a novel dual-branch Mamba-based model. Our architecture utilizes Mamba's linear-time modeling to efficiently learn distinct spatial and spectral features in specialized branches, which are then integrated by a dynamic gated fusion mechanism to enhance anomaly localization. Across fourteen benchmark HSI datasets, our proposed DMS2F-HAD not only achieves a state-of-the-art average AUC of 98.78%, but also demonstrates superior efficiency with an inference speed 4.6 times faster than comparable deep learning methods. The results highlight DMS2FHAD's strong generalization and scalability, positioning it as a strong candidate for practical HAD applications.
Paper Structure (25 sections, 2 equations, 5 figures, 5 tables)

This paper contains 25 sections, 2 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Background and Related Works
  3. Hyperspectral Anomaly Detection
  4. State Space Models
  5. Mamba Models
  6. The Proposed Method
  7. Overall Framework
  8. Data Pre-Processing
  9. Dual-branch Encoder Block
  10. Spatial Branch:
  11. Spectral Branch:
  12. Adaptive Gated Fusion
  13. Decoder and Reconstruction
  14. Anomaly Detection
  15. Experiments and Analysis
  16. ...and 10 more sections

Figures (5)

  • Figure 1: The accuracy-efficiency trade-off on the AVIRIS-2 dataset, where circle size reflects model parameters. Our DMS2F-HAD clearly achieves the best balance, delivering the highest AUC with minimal complexity.
  • Figure 2: (a) Overall architecture of DMS2F-HAD framework; (b) Mamba block; and (c) SS Decoder Block
  • Figure 3: Colour Anomaly maps of different HAD methods on six datasets
  • Figure 4: Box-whisker plots of anomaly scores of different HAD methods on six datasets.
  • Figure 5: 2D ROC curves of seven different methods on six datasets.