Benchmarking Deep Learning and Statistical Target Detection Methods for PFM-1 Landmine Detection in UAV Hyperspectral Imagery
This repository contains the code and annotations for our IGARSS 2026 paper on "Benchmarking Deep Learning and Statistical Target Detection Methods for PFM-1 Landmine Detection in UAV Hyperspectral Imagery" by Sagar Lekhak, Prasanna Reddy Pulakurthi, Ramesh Bhatta, and Emmett J. Ientilucci.
This dataset 1 contains Visible and Near-Infrared (VNIR) Hyperspectral Imaging (HSI) data prepared for the IEEE International Geoscience and Remote Sensing Symposium (IGARSS) 2026.
This specific version is a refined subset of the original benchmark dataset 2. While the original release provided full radiance cubes, broad GCP/AeroPoint data, and reference ground spectra of all the targets, this version focuses on a spatially subsetted region containing only PFM-1 landmine targets and includes high-accuracy, pixel-wise binary ground truth masks.
The data was collected over a controlled test field seeded with 143 realistic surrogate landmine and UXO targets (surface, partially buried, and fully buried). Data acquisition was performed using a Headwall Nano-Hyperspec® sensor mounted on a multi-sensor UAV platform flown at an altitude of ~20.6 m 2.
For more details regarding data acquisition and preprocessing, go to the original paper 2.
- Sensor: [Headwall Nano-Hyperspec®]
- Spectral Range: [398–1002 nm]
- Number of Bands: [270 bands]
- Approximate GSD: [Approx. 1.29 cm]
# Clone the repository
git clone https://github.com/PrasannaPulakurthi/pfm1-hsi-benchmark.git
cd pfm1-hsi-benchmark
# Create a virtual environment
python -m venv .venv
source .venv/bin/activate
# Install dependencies
pip install -r requirements.txtfrom huggingface_hub import snapshot_download
snapshot_download(repo_id="SagarLekhak/pfm1-landmine-uav-vnir-hsi-IGARSS-2026", repo_type="dataset", local_dir="./data")# Full Region evaluation
jupyter notebook Full_Region.ipynb
# PFM-1 Region evaluation
jupyter notebook PFM-1_Region.ipynb
# Independent Test Region evaluation
jupyter notebook FL_Test_Region.ipynbjupyter notebook Visualize_Data.ipynbVisualize the hyperspectral dataset, ground truth masks, and spectral signatures.
| Method | Average Precision (AP) | ROC-AUC |
|---|---|---|
| SAM | 0.144 | 0.797 |
| MF | 0.358 | 0.982 |
| ACE | 0.691 | 0.989 |
| CEM | 0.589 | 0.983 |
| Spectral-NN | 0.814 | 0.982 |
- ROC-AUC alone is misleading: All methods achieve high AUC (>0.98) but vastly different precision-recall performance
- Spectral-NN achieves best AP: Outperforms all classical methods with AP=0.814
- Scene composition matters: Performance varies significantly between Full Region, PFM-1 Region, and Test Region
- Precision-focused evaluation is critical: For rare-target detection with extreme class imbalance
Precision-Recall curves
Log-scale ROC curves revealing Spectral-NN's superior performance at operationally critical low false-positive rates
If you use this dataset, please cite the specific IGARSS 2026 work 1 and the original benchmark 2:
@misc{lekhak2026benchmarkingdeeplearningstatistical,
title={Benchmarking Deep Learning and Statistical Target Detection Methods for PFM-1 Landmine Detection in UAV Hyperspectral Imagery},
author={Sagar Lekhak and Prasanna Reddy Pulakurthi and Ramesh Bhatta and Emmett J. Ientilucci},
year={2026},
eprint={2602.10434},
archivePrefix={arXiv},
primaryClass={eess.IV},
url={https://arxiv.org/abs/2602.10434},
}[2] Original Benchmark Dataset:
@misc{lekhak2026uavbasedvnirhyperspectralbenchmark,
title={A UAV-Based VNIR Hyperspectral Benchmark Dataset for Landmine and UXO Detection},
author={Sagar Lekhak and Emmett J. Ientilucci and Jasper Baur and Susmita Ghosh},
year={2026},
eprint={2510.02700},
archivePrefix={arXiv},
primaryClass={eess.IV},
url={https://arxiv.org/abs/2510.02700},
}This project is licensed under the MIT License - see the LICENSE file for details.