Attention and Edge-Aware Band Selection for Efficient Hyperspectral Classification of Burned Vegetation

This repository provides the official implementation of the methods described in our MLSP 2025 paper:
"Attention and Edge-Aware Band Selection for Efficient Hyperspectral Classification of Burned Vegetation".
We investigate two complementary hyperspectral band selection methods—Spatial-Spectral Edge Preservation (SSEP) and Spectral-Redundancy Penalized Attention Ranking (SRPA)—for classifying vegetation in post-burn environments using VNIR hyperspectral imagery.

---

🔬 Overview

Effective post-burn vegetation classification is critical for ecological recovery and wildfire risk assessment.
Hyperspectral imaging (HSI) offers rich spectral information but suffers from high dimensionality.
This repository implements: - SSEP: An unsupervised, edge-aware method aligning spectral and label-derived edges. - SRPA: A supervised attention-based method with redundancy penalization for discriminative band selection.

Both methods are evaluated using: - Random Forest (RF): Classical spectral-only classifier. - Lightweight 3D CNN: Spatial-spectral deep learning model optimized for efficiency.

---

📂 Repository Structure

SF_Prescribed_Fire_HSI/ │ ├── data/ # VNIR hyperspectral cube (ENVI format) and label masks ├── preprocessing/ # Scripts for patch extraction, normalization, and labeling ├── band_selection/ # Implementations of SSEP and SRPA algorithms ├── models/ # 3D CNN architecture and training scripts ├── rf_baseline/ # Random Forest training pipeline ├── visualization/ # Band score plots and classification results └── utils/ # Helper functions (e.g., Dice score, redundancy computation)

---

🛠 Installation

1. Clone the repository: bash git clone https://github.com/BMW-lab-MSU/SF_Prescibed_Fire_HSI.git cd SF_Prescibed_Fire_HSI
2. Create and activate a virtual environment: bash python3 -m venv venv source venv/bin/activate # Linux/Mac venv\Scripts\activate # Windows
3. Install dependencies: bash pip install -r requirements.txt

---

📊 Dataset

The hyperspectral dataset was collected over a thin-burn plot (330m × 300m) in the Lubrecht Experimental Forest (Montana, USA) using a Headwall VNIR sensor (273 bands, 400–1000 nm).
- Labels include Tree, Grass, and Soil, manually annotated from RGB composites derived from hyperspectral bands.

Note: Dataset access may require permission. Contact the maintainers for availability.

---

🚀 Usage

1️⃣ Preprocessing & Patch Extraction

bash python preprocessing/extract_patches.py --cube_path data/VNIR.hdr --labels data/labels.png --patch_size 50 --stride 25

2️⃣ Run Band Selection

• SSEP: bash python band_selection/ssep.py --cube_path data/VNIR.hdr --labels data/labels.png --top_k 50
• SRPA: bash python band_selection/srpa.py --patch_dir data/patches --labels data/patch_labels.npy --top_k 50 --lambda_penalty 0.2

3️⃣ Train Models

• Random Forest (RF): bash python rf_baseline/train_rf.py --bands_selected outputs/ssep_top50.npy
• 3D CNN: bash python models/train_3dcnn.py --bands_selected outputs/srpa_top50.npy

---

📈 Results (Key Findings)

• SRPA outperformed SSEP across all metrics:
  • Best Accuracy: 93.89% (SRPA + 3D CNN, Top-50 bands)
    
  • Best F1 Score: 48.31% (SRPA + 3D CNN)
    
• SSEP achieved competitive results in low-dimensional regimes (Top-10 bands).
  
• SRPA’s attention-guided band selection better complements spatial-spectral deep learning models.

---

📊 Visualizations

• Band importance plots (SSEP & SRPA)
• Classification accuracy vs. Top-k bands
• Spectral signature overlays for Tree and Grass
  

---

🔗 Citation

If you use this repository or its methods, please cite, will update correct citations soon: @inproceedings{mlsp2025_hsi_burn, title={Attention and Edge-Aware Band Selection for Efficient Hyperspectral Classification of Burned Vegetation}, booktitle={IEEE International Workshop on Machine Learning for Signal Processing (MLSP)}, year={2025}, author={Mahmad Isaq Karankot} }

---

🤝 Acknowledgements

This work is supported by NSF EPSCoR (OIA-2242802) as part of the SMART FIRES project at Montana State University.

🚀 Quick Start Guide

Follow these steps to quickly run the pipeline and reproduce results:

1️⃣ Download or Prepare Dataset

• Place the VNIR hyperspectral cube (VNIR.hdr and VNIR.raw) in the data/ directory.
  
• Place the manually annotated label mask (labels.png) in the same folder.

2️⃣ Extract Patches

bash python preprocessing/extract_patches.py --cube_path data/VNIR.hdr --labels data/labels.png --patch_size 50 --stride 25 ✅ Output: Extracted patches saved in data/patches/ and labels in data/patch_labels.npy.

3️⃣ Run Band Selection

• SSEP (Edge-Aware): bash python band_selection/ssep.py --cube_path data/VNIR.hdr --labels data/labels.png --top_k 50 ✅ Output: outputs/ssep_top50.npy (Top-50 bands).

• SRPA (Attention-Based): bash python band_selection/srpa.py --patch_dir data/patches --labels data/patch_labels.npy --top_k 50 --lambda_penalty 0.2 ✅ Output: outputs/srpa_top50.npy (Top-50 bands).

4️⃣ Train Models

• Random Forest: bash python rf_baseline/train_rf.py --bands_selected outputs/ssep_top50.npy
• 3D CNN: bash python models/train_3dcnn.py --bands_selected outputs/srpa_top50.npy ✅ Output: Validation accuracy and F1 scores logged in outputs/logs/.

5️⃣ Visualize Results

bash python visualization/plot_band_scores.py --method ssep --scores outputs/ssep_scores.npy python visualization/plot_band_scores.py --method srpa --scores outputs/srpa_scores.npy ✅ Output: Band importance plots and classification result charts in outputs/figures/.

---

🔥 Sample Results

• SRPA + 3D CNN (Top-50 bands):
  
  • Accuracy: 93.89%
  • Macro F1 Score: 48.31%
    
• SSEP + RF (Top-50 bands):
  
  • Accuracy: 79.98%
  • Macro F1 Score: 71.89%
    

---