Official implementation of “Supervised Embedded Methods for Hyperspectral Band Selection”

Yaniv Zimmer, Ofir Lindenbaum, and Oren Glickman — ECAI (2025)

Read the paper

ArXiv

CHBS: Concrete Hyperspectral Band Selection

---

🌈 Overview

CHBS (Concrete Hyperspectral Band Selection) is a supervised, embedded method for task-specific hyperspectral band selection.

Unlike traditional preprocessing-based methods, CHBS integrates band selection directly into a deep learning model through a Concrete Selector Layer using Gumbel–Softmax reparameterization.

This design enables differentiable, end-to-end training that jointly optimizes both band selection and downstream task performance, making CHBS ideal for real-time or resource-constrained environments such as autonomous driving and environmental sensing.

---

🔍 Key Features

• 🚀 Embedded Band Selection: No separate preprocessing — fully end-to-end.
• 🧩 Concrete Selector Layer: Differentiable band selection via Gumbel–Softmax.
• 🎯 Segmented Xavier Initialization: Ensures spectral diversity and stable convergence.
• 🧠 Task-Aware Optimization: Selects bands that improve model accuracy directly.
• 📈 State-of-the-Art Results: Outperforms prior methods on multiple HSI benchmarks.

---

🧠 Method Summary

CHBS learns k informative spectral bands by training a learnable logits matrix L via a Gumbel–Softmax distribution:

$$ M_{i,j} = \frac{\exp((L_{i,j} + G_{i,j}) / \tau)}{\sum_r \exp((L_{i,r} + G_{i,r}) / \tau)} $$

This produces a soft, differentiable band-selection mask that becomes nearly discrete as the temperature τ decreases during training.

---

⚙️ Installation

git clone https://github.com/YanivZimmer/CHBS.git cd CHBS pip install -r requirements.txt

🧩 Acknowledgements

This work was partially supported by the Chief Scientist of the Israeli Ministry of Agriculture, grant number 12-03-0010.

📚 Citation

please cite:

@article{zimmer2024embedded,
  title={Embedded hyperspectral band selection with adaptive optimization for image semantic segmentation},
  author={Zimmer, Yaniv and Glickman, Oren},
  journal={arXiv preprint arXiv:2401.11420},
  year={2024}
}