DMS2F-HAD: Dual Mamba Spectral-Spatial Fusion for Hyperspectral Anomaly Detection

DMS2F-HAD is a deep learning framework for hyperspectral anomaly detection using a dual-branch Mamba architecture that adaptively fuses spectral and spatial features with gated blocks.

Features

• Dual-branch Mamba architecture (spectral + spatial)
• Gated fusion for enhanced anomaly detection
• Patch-wise image preprocessing with random masking
• Built-in evaluation and ROC visualization tools
• PyTorch-based, modular, and easily extendable

Installation

# 1. Create environment
conda create -n mamba_env python=3.12 -y
conda activate mamba_env

# 2. Install dependencies
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
pip3 install timm==0.8.15dev0 mmselfsup pandas transformers openpyxl imgaug numba numpy tensorboard fvcore Ninja mmdet==2.25.3
pip3 install --upgrade protobuf==3.20.1 scikit-image faiss-gpu
pip3 install adeval fastprogress geomloss FrEIA mamba_ssm triton causal_conv1d numpy-hilbert-curve pyzorder
conda install -c conda-forge accimage
pip install -r requirements.txt

## 3. Download folder

Download the code folder
cd DMS2FHAD

Project Structure:

📁 DMS2F-HAD/
├── data/                          
├── artifacts/
│   ├── models/
│   └── results/
├── experiments/
│   ├── ablation/
├── docs/
├── src/                
│   ├── block_utils.py
│   ├── mask.py
│   ├── train.py 
│   ├── data_loader.py.py    
│   ├── mamba_simple.py 
│   ├── model.py     
│   └── analyze_residual.py 
├── main.py
├── requirements.txt
└── README.md

Prepare Dataset:

Datasets are available in DMS2F-HAD/data.

-- cri.mat
-- aviris1.mat
-- aviris2.mat
-- SanDiego.mat

Usage

CUDA_VISIBLE_DEVICES=0 
python main.py
python src/analyze_residual.py

Demo

python src/analyze_residual.py

Results are saved in artifacts/results/current.

Software Requirements

• OS: Ubuntu 20+, macOS, or Windows 10+
• Python: 3.12.9
• PyTorch: ≥ 2.5.1 with CUDA 12.1
• GPU: Recommended (NVIDIA,greater than A100 with CUDA installed)
• Package Manager: Anaconda (recommended)