Benchmarking Deep Learning and Statistical Target Detection Methods for PFM-1 Landmine Detection in UAV Hyperspectral Imagery

TL;DR

This work tackles the lack of standardized benchmarks for UAV-based hyperspectral landmine detection by systematically evaluating four classical spectral detectors (SAM, MF, ACE, CEM) alongside a lightweight Spectral Neural Network on a VNIR UAV dataset of PFM-1 targets. The methodology includes careful preprocessing, ground-truth annotation, and analysis using ROC-AUC and AP to capture performance under severe class imbalance. Key findings show ACE delivers the strongest ROC-AUC, while the Spectral-NN excels in precision–recall, underscoring the need for precision-focused metrics and scene-aware benchmarking in operational settings. The study also provides pixel-level ground-truth masks to enable reproducible research and motivates future work in uncertainty-aware and spatial–spectral models, as well as multi-sensor fusion for comprehensive surface and shallowly buried mine detection.

Abstract

In recent years, unmanned aerial vehicles (UAVs) equipped with imaging sensors and automated processing algorithms have emerged as a promising tool to accelerate large-area surveys while reducing risk to human operators. Although hyperspectral imaging (HSI) enables material discrimination using spectral signatures, standardized benchmarks for UAV-based landmine detection remain scarce. In this work, we present a systematic benchmark of four classical statistical detection algorithms, including Spectral Angle Mapper (SAM), Matched Filter (MF), Adaptive Cosine Estimator (ACE), and Constrained Energy Minimization (CEM), alongside a proposed lightweight Spectral Neural Network utilizing Parametric Mish activations for PFM-1 landmine detection. We also release pixel-level binary ground truth masks (target/background) to enable standardized, reproducible evaluation. Evaluations were conducted on inert PFM-1 targets across multiple scene crops using a recently released VNIR hyperspectral dataset. Metrics such as receiver operating characteristic (ROC) curve, area under the curve (AUC), precision-recall (PR) curve, and average precision (AP) were used. While all methods achieve high ROC-AUC on an independent test set, the ACE method observes the highest AUC of 0.989. However, because target pixels are extremely sparse relative to background, ROC-AUC alone can be misleading; under precision-focused evaluation (PR and AP), the Spectral-NN outperforms classical detectors, achieving the highest AP. These results emphasize the need for precision-focused evaluation, scene-aware benchmarking, and learning-based spectral models for reliable UAV-based hyperspectral landmine detection. The code and pixel-level annotations will be released.

Figures (5)

• Figure 1: Overview of PFM-1 mines (highlighted in red) and other target types (Image source baur2023accessible).
• Figure 2: Cropped regions used in this study from the original hyperspectral dataset lekhak2025uav: (top) Full Region, containing all mine locations; (middle) PFM-1 Region, showing only PFM-1 targets; (bottom) Ground truth binary mask of the PFM-1 Region, with training set containing the first five PFM-1 targets (left) and the remaining two targets used for testing (right).
• Figure 3: ROC curves and corresponding AUCs for all algorithms in the Test Region: a) linear scale and b) logarithmic scale.
• Figure 4: Precision-recall curve and corresponding APs for a) Full Region (left), b) PFM-1 Region (middle), and Test Region (right).
• Figure 5: Aeropoints in the scene appearing as the biggest false alarms for PFM-1 mines even at higher threshold values for detection scores in classical statistical methods in Full Region. Presented here is an example case for a representative thresholds (0.77 -- 0.85) in MF detection scores.