Description: This repository provides an end‑to‑end, research‑grade workflow for quantitative mineral mapping from PRISMA (PRecursore IperSpettrale della Missione Applicativa) hyperspectral imagery. It spans every processing stage, from physics‑informed spectral‑library augmentation through adaptive endmember detection to a compact 3‑D Convolutional Neural Network (CNN), and converts raw Level‑2D radiances into georeferenced mineral‐class and abundance maps that faithfully reproduce the results in our ISPRS JPRS article on the McMurdo Dry Valleys, Antarctica.

Scientific scope and capabilities

• Sensor coverage – validated on PRISMA (400–2500 nm, 30 m, 239 bands) but configurable for EMIT, AVIRIS‑NG, HyMap, EnMAP, DESIS, or laboratory cubes with minimal code changes.
• Spectral‑library augmentation – stochastic offset perturbation, sub‑pixel mixture synthesis, and band‑specific noise injection generate >10 000 physically plausible reflectance spectra per class, mitigating class imbalance and enhancing model generalisation under novel illumination conditions.
• Adaptive VCA (AVCA) – automatically estimates the intrinsic dimensionality R via Harsanyi’s virtual dimensionality (VD) test and employs SNR‑aware orthogonal projections to derive robust endmembers even under low contrast (SNR < 30 dB) and ice‑dust contamination.
• 3‑D CNN classifier + soft‑unmixer – an 8‑layer, 1.25 M‑parameter network with residual shortcuts that jointly learns spatial–spectral patterns in 27 × 27 × R hyper‑patches, outputting both crisp categorical maps and continuous abundance cubes through a dual‑branch Dice‑Focal loss architecture.
• Reproducibility & geospatial fidelity – deterministic seeds, YAML experiment manifests, conda‑lock environment capture, and automatic propagation of ENVI map‑info (UTM 58S, WGS‑84) to all outputs guarantee that published metrics can be reproduced to within ±1 % overall accuracy and ±0.004 RMSE.

Potential applications include geothermal alteration mapping, critical‑mineral prospecting, ice‑sheet sediment characterisation, and autonomous rover vision. Researchers can rapidly retarget the CNN to new sensors or lithologies by replacing the spectral libraries and adjusting a single config.yaml file.

Revealing Critical Mineralogical Insights in Extreme Environments Using Deep‑Learning on PRISMA Hyperspectral Imagery – Dry Valleys, South Victoria Land, Antarctica)

Authors: Jabar Habashi · Amin Beiranvand Pour · Aidy M Muslim · Ali Moradi Afrapoli · Jong Kuk Hong · Yongcheol Park · Alireza Almasi · Laura Crispini · Mazlan Hashim · Milad Bagheri Journal: ISPRS Journal of Photogrammetry and Remote Sensing 228, 83–121. Status / DOI:10.1016/j.isprsjprs.2025.07.005*

1. Overview

2. Repository Layout

3. Quick Start

   1. Environment
   2. Installation

4. Workflow

   1. 1️⃣ Spectral‑Library Augmentation
   2. 2️⃣ Adaptive VCA Endmember Extraction
   3. 3️⃣ 3‑D CNN Classification & Abundance Mapping

5. Expected Inputs & Directory Structure

6. Outputs

7. Reproducibility Checklist

8. Citation & License

9. Contact & Support

This repository hosts three standalone yet interoperable Python scripts that together comprise the complete processing chain used in the above‑cited article to augment spectral libraries, extract endmembers and unmix hyperspectral PRISMA imagery with a deep 3‑D Convolutional Neural Network (CNN). Each module can be executed independently, but maximum performance is obtained when they are run sequentially:

All three scripts are released under the MIT License with the requirement that users cite the accompanying article if any part of the code is employed in academic or commercial work.

Important: The scripts contain placeholder paths (marked by comments such as # import RS data path) that must be updated to match your local directory structure before execution.

📂 hyperspectral‑mineral‑mapping/
├─ Spectral Augmentation.py           # Spectral offset augmentation
├─ AVCA.py                            # Adaptive Vertex Component Analysis
├─ 3D-CNN.py                          # 3‑D CNN classifier & mapper
├─ examples/
│   ├─ spectra/                       # Sample spectral libraries (.sli / .hdr)
│   ├─ prisma_scene/                  # Sample PRISMA cube (.dat / .hdr)
│   └─ mask/                          # Example binary mask
├─ docs/                              # Supplementary figures & slides
└─ README.md                          # You are here ✅

Feel free to reorganise as long as the internal path variables in the scripts are updated accordingly.

# Clone the repo
$ git clone https://github.com/<YourUser>/hyperspectral‑mineral‑mapping.git
$ cd hyperspectral‑mineral‑mapping

# (Recommended) Create a dedicated environment
$ conda create -n hypermap python=3.10
$ conda activate hypermap

# Install core dependencies
$ pip install -r requirements.txt

Minimum tested package versions (adjust as needed):

• numpy ≥ 1.25
• rasterio ≥ 1.3
• spectral ≥ 0.23
• tensorflow ≥ 2.15 (with GPU support preferred)
• scikit‑learn ≥ 1.4
• seaborn ≥ 0.13
• matplotlib ≥ 3.9

Tip: GPU acceleration (CUDA 11+) reduces CNN training time by ~10×.

No additional compilation is required; the scripts are pure Python. After installing the dependencies, you are ready to run the pipeline.

The following diagram summarises the recommended execution order:

           ┌──────────────────┐      ┌─────────────────┐      ┌─────────────────┐
           │   Spectral SLI   │      │   PRISMA Cube   │      │   Binary Mask   │
           └────────┬─────────┘      └────────┬────────┘      └────────┬────────┘
                    │                         │                        │
                    ▼                         │                        │
       Spectral Augmentation.py               │                        │
           (augmented *.txt)                  │                        │
                    │                         │                        │
                    │                         ▼                        │
                    │                      AVCA.py                     │
                    │                 (endmember matrix)               │
                    │                         │                        │
                    └──────────────┬──────────┴──────────┬─────────────┘
                                   ▼                     ▼
                                3D-CNN.py  –––▶  Classified Map &
                                         Abundance Cubes

Input requirement: The spectral profiles you supply **must be exported from **ENVI Classic as paired .hdr files. Libraries saved in newer ENVI variants can lack essential wavelength metadata, causing the augmentation script to fail.

$ python Spectral Augmentation.py \
      --input r"C:\path\to\ASCL_library" \
      --output Augmented Spectral Libraries.txt \
      --iterations 30 --offset 0.01

• Input: ENVI‑classic spectral library (*.sli) of mineral reflectances.
• Output: Vegetation.txt – n × (m·iter) matrix where n is the number of wavelengths and m is the original number of spectra.

Modify initial_offset, increment and num_iterations in‑code or expose them via CLI flags to tailor augmentation strength.

$ python AVCA_for_shairing.py \
      --raster r"C:\path\prisma\scene" \
      --mask   r"C:\path\mask" \
      --hdr    r"C:\path\prisma\scene.hdr" \
      --R 15

• Input: PRISMA hyperspectral cube (*.dat/*.hdr) & optional binary ROI mask.
• Output: AVCA_results.txt – wavelengths + R endmember spectra.

The script internally estimates SNR and switches between projective & subspace projections as per [Nascimento & Dias, 2005].

$ python CNN_for_shaire.py \
      --rs       r"C:\path\prisma\scene" \
      --hdr      r"C:\path\prisma\scene.hdr" \
      --mask     r"C:\path\mask" \
      --hdr      r"C:\path\mask.hdr" \
      --sli_dir  r"C:\path\spectral_libraries" \
      --epochs   300 --batch_size 64

• Input:

  • PRISMA scene & HDR
  • Binary mask (1 = valid pixel)
  • Folder of augmented spectral libraries

• Outputs:

  • classified.png – mineral class map
  • abundance.hdr/dat – BSQ cube, one band per class
  • training_log.csv – accuracy / loss history

Early Stopping: Training halts when both training & validation accuracy ≥ 0.993 and loss ≤ 0.11 (modifiable via CustomEarlyStopping).

├─ data/
│  ├─ prisma_scene.dat / .hdr
│  ├─ mask.dat / .hdr
│  └─ spectral_libraries/
│     ├─ Mineral_1.sli / .hdr
│     ├─ Mineral_2.sli / .hdr
│     └─ ...

Ensure consistent interleave (bsq) and byte order across all ENVI files.

All outputs inherit geospatial metadata (map info & projection) from the input HDR to ensure GIS compatibility.

This code is released under the permissive MIT License (see headers in each script). If you use any part of this repository, please cite our article:

Habashi, J., Pour, A.B., Muslim, A.M., Afrapoli, A.M., Hong, J.K., Park, Y., Almasi, A., Crispini, L., Hashim, M., Bagheri, M., 2025. Revealing critical mineralogical insights in extreme environments using deep learning technique on hyperspectral PRISMA satellite imagery: Dry Valleys, South Victoria Land, Antarctica. ISPRS Journal of Photogrammetry and Remote Sensing 228, 83–121. https://doi.org/10.1016/j.isprsjprs.2025.07.005

Silva, J. M. P. Nascimento & J. M. B. Dias, "Vertex Component Analysis: a Fast Algorithm to Unmix Hyperspectral Data," IEEE TGRS, 43(4):898–910, 2005.

• Lead Developer: Jabar Habashi – j_habashi98 [at] sut.ac.ir
• Issues and pull requests are very welcome! If you discover bugs or have suggestions for improvement, please open an issue and fill out the template.