UAV-Based Hyperspectral Imaging Dataset of Pearl Millet Water Stress

This repository contains the dataset description, preprocessing workflow, benchmark model implementations, and performance results as described in the IEEE paper:

"UAV-Borne Hyperspectral Imaging Dataset of Pearl Millet Canopy Water Stress"
IEEE (published in ICA 2024)

Dataset Overview

• Crop: Pearl Millet
• Sensor: Resonon Pika-L HSI camera (282 bands, 400–1000 nm)
• Platform: DJI Matrice-600 Pro UAV
• Water Treatments: Well-Watered (WW), Water-Stressed (WS)
• Stress Levels: WS1 (12 days after irrigation stopped), WS2 (18 days)
• Genotypes: 350 varieties × (4 WW + 5 WS) replications
• Samples: 25×25×282 HS image patches

| Dataset | Total Samples | Train Samples | Test Samples | |---------|----------------|----------------|----------------| | WW vs WS1 | 10,399 | 4662 | 5737 | | WW vs WS2 | 5718 | 2667 | 3051 |

Dataset Access (via TiHAN Server)

Due to data size and controlled access requirements, the dataset is hosted on the TiHAN server.

Navigate to Agricultural Dataset => Hyperspectral Imaging => Request access => Fill the form

📝 Request Dataset Access

Access will be granted upon filling a form explaining your research purpose.

Dataset Creation Workflow

Steps followed as per the paper:

1. Radiometric Calibration – Convert raw DN to radiance and reflectance.
2. Geometric Rectification – Remove motion distortion, geo-align plots.
3. Image Quality Check – NIQE used for filtering.
4. Denoising – Using FastHyDe for spectral consistency.
5. Patch Extraction – NDVI > 0.7, TCARI > 1300 based segmentation.

Benchmark Models

The following models are implemented for classification:

• SVM (Spectral classifier)
• 3D-CNN
• 2D+1D CNN
• 3D-CAE
• ViT (Vision Transformer)
• SpectralFormer

You’ll find implementation scripts under code/classification/.

Performance Metrics

| Model | WW-WS1 OA (%) | WW-WS2 OA (%) | |-------|----------------|----------------| | SVM | 77.87 | 79.82 | | 3D-CAE | 81.31 | 92.19 | | 2D+1D CNN | 73.01 | 88.59 | | 3D-CNN | 82.07 | 88.52 | | ViT | 81.26 | 89.92 | | SpectralFormer | 81.51 | 93.88 |

Folder Structure

bash HSI-PearlMillet-WaterStress/ ├── dataset/ # Dataset metadata, access instructions │ └── README.md ├── code/ # All model and preprocessing code │ ├── preprocessing/ │ ├── classification/ │ └── utils/ ├── models/ # Trained models ├── results/ # Graphs, OA scores, confusion matrices ├── LICENSE ├── CITATION.bib └── README.md

Citation

bibtex @inproceedings{sankararao2025hsi, author = {Sankararao, Adduru U. G. and Kiran, Sai and Rajalakshmi, P. and Choudhary, Sunita}, title = {UAV-Borne Hyperspectral Imaging Dataset of Pearl Millet Canopy Water Stress}, booktitle = {International Conference on Advances in Computing}, year = {2025}, doi = {10.1007/978-3-031-74440-2_19} }

Contributors

• Dr. Adduru U.G. Sankararao – IIT Hyderabad
• Mr. Sai Kiran – IIT Hyderabad
• Prof. P. Rajalakshmi – Research Advisor
• Dr. Sunita Choudhary – Scientist, ICRISAT

Notes

• Pretrained models may be shared later depending on approval.
• Suggestions/Issues welcome via GitHub Issues.

© 2025 IIT Hyderabad & ICRISAT – All Rights Reserved.