Source-linked AI summary
UAV Thermal Imagery for Inert Ordnance Screening: Multi Campaign Dataset Development,Object Detection, and Practical Recommendations
Chad Melton, PhD., Annabelle Kelton
TL;DR
UXO contamination creates a need for safer, better-supported remote screening, while prior work offers limited end-to-end evidence from field collection through operational guidance. This study builds and labels a multi-campaign UAV thermal dataset, evaluates two candidate detectors, and derives practical recommendations. Both models performed best in the 15 m implementation, while outputs remain intended for analyst-supported prioritization rather than standalone clearance.
Problem
Relatively few prior studies describe one sequence spanning field campaigns, labeled dataset development, automated screening evaluation, and practical guidance for UXO response.
Method
The study collected and labeled multi-condition UAV thermal imagery, constructed datasets, trained YOLOV11l and RT-DETR-R50, and evaluated analyst-supported screening workflows.
Results
5,855 thermal image label pairs were collected, and both YOLOV11l and RT-DETR-R50 achieved their strongest reported results with the 15 m dataset.
Takeaways & Limitations
Thermal and RGB collection, varied backgrounds and surfaces, post-solar-exposure scheduling, representative local calibration, and qualified human review are recommended for screening and follow-on assessment.
Takeaways & Limitations
The study did not benchmark the detectors directly, and geospatial grouping, realistic false-alarm testing, and operational clearance validation remain incomplete.
Abstract
from arXiv · showhide
Unexploded ordnance (UXO) continues to restrict civilian access, agricultural activity, infrastructure recovery, and environmental remediation in contaminated areas around the world. This study created a multi campaign UAV thermal image data set of inert ordnance, developed a labeled image set from collected imagery, tested object detection models, and identified practical considerations for humanitarian mine action and demining applications. Data were collected during four field campaigns in Tennessee under summer and winter conditions using inert mines, munitions, and other ordnance placed in short grass, tall vegetation, gravel, mulch, rock, compost, and compacted surfaces. Thermal imagery was collected under flight altitutes of 33 m and 15 m. The final source inventory contained 5,855 thermal image label pairs, including 918 positive images and 4,937 background images. After retaining all positive images and downsampling background images, the 33 m dataset contained 420 training and 106 validation images, while the 15 m dataset contained 629 training and 157 validation images. YOLOV11l and RT-DETR-R50 algorithms were trained and evaluated to develop an automated candidate detection model. Practical recommendations include collecting thermal and RGB imagery together, incorporating varied surfaces and background only imagery, considering periods following changes in solar exposure, balancing survey coverage against target pixel representation, calibrating models with representative local data, and retaining qualified human review. The intended use is screening and prioritization for follow on technical survey or EOD assessment, and not a standalone clearance.
Background
UAV thermal imagery offers a remote complement to ground-based UXO investigation, but prior studies provide limited end-to-end evidence linking field collection, dataset construction, automated screening, and operational guidance.
- Ground-based UXO investigation can expose personnel, civilians, equipment, and soil to detonation-related risks.
- Thermal imaging complements material-specific sensors by recording surface radiation and revealing temperature differences from material properties, solar loading, and heat transfer.Buried-object detection depends on whether ordnance produces a measurable surface temperature anomaly.
- Earlier deep-learning work reported precision higher than 98% for YOLOv5 UXO detection, while noting that its small dataset limited reproducibility with larger datasets.
- Prior UAV thermal studies detected some plastic, ceramic, and wood-encased ordnance but reported constraints involving altitude, timing, vegetation, burial, weather, and material type.One Burkina Faso study reported a highest detectability of 75 percent and did not detect metal landmines.
- Relatively few studies combine field campaign design, labeled dataset creation, automated screening evaluation, and practical guidance in one workflow.
- This study addresses that gap by creating a multi-campaign UAV thermal dataset, documenting labeling and dataset preparation, evaluating candidate detection, and developing demining guidance.
Methodology
The study links multi-condition UAV thermal/RGB collection with manual labeling, dataset preparation, object-detection evaluation, analyst review, and prioritization for follow-on investigation.
- Methodology: The methodology proceeds from mission planning and environmental scheduling through thermal/RGB collection, quality control, annotation, dataset construction, model validation, analyst review, and follow-on prioritization.
- Evaluation Scope and Interpretation: The experiments directly evaluated field collection, annotation, dataset development, and image-level object detection under controlled conditions.
- Field Campaigns and Experimental Design: Four Tennessee campaigns varied season, vegetation, surface material, solar loading, and flight altitude, including three campaigns at 33 m and one at approximately 15 m.
- Field Campaigns and Experimental Design: Targets were positioned across tall and short vegetation and varied surfaces to create partial-obstruction and background variability rather than reproduce a complete operational minefield.
- Aircraft and Sensor Payloads: The study used two UAV platforms, with the Skydio X10 combining high-resolution RGB cameras and a radiometric 640 × 512 thermal sensor for same-flight collection.
- Data Labeling and Dataset Preparation: Thermal imagery was manually reviewed in CVAT, visible ordnance was bounding-box labeled, and all target categories were remapped to a single uxo class.
- Evaluation Scope and Interpretation: YOLOV11l and RT-DETR-R50 were evaluated within their native training and evaluation procedures, so the study was not designed as a controlled architecture benchmark.
Results
The study produced labeled thermal datasets from four campaigns and found higher validation performance at 15 m than 33 m for both evaluated detectors, while dataset composition limits operational interpretation.
- Dataset outcomes: 5,855 thermal image label pairs comprised 918 positive images and 4,937 background images across four campaigns.The selected datasets contained 526 images at 33 m and 786 images at 15 m.
- Dataset outcomes: At 33 m, the selected dataset contained 420 training and 106 validation images, while the 15 m dataset contained 629 training and 157 validation images.The validation subsets included 176 annotated instances at 33 m and 192 at 15 m.
- Object detection results: For YOLOV11l, 15 m increased precision from 0.890 to 0.925, recall from 0.828 to 0.833, F1 from 0.858 to 0.877, mAP50 from 0.865 to 0.905, and mAP50-95 from 0.488 to 0.664.The largest change was an mAP50-95 increase of 0.176.
- Object detection results: For RT-DETR-R50, 15 m increased precision from 0.891 to 0.911, recall from 0.699 to 0.750, F1 from 0.783 to 0.823, mAP50 from 0.644 to 0.740, and mAP50-95 from 0.316 to 0.511.The largest difference was an mAP50-95 increase of 0.195.
- Interpretation boundary: The validation subsets used approximately 70 percent positive and 30 percent background images, so precision should not be interpreted as an operational false alarm rate.The two model families also used different evaluator and confidence-calibration configurations, and the RT-DETR-R50 regimes did not use identical input-resolution policies.
Discussion
Thermal detection performance varied with acquisition conditions, while operational use depends on timing, altitude, model calibration, and qualified human review. The authors frame the detectors as screening tools for prioritizing follow-on investigation rather than standalone clearance systems.
- Field Observations and Acquisition Timing: 30 to 60 minutes after sunrise or sunset, and immediately after cloud transitions, were preliminary periods when targets appeared most distinguishable.The authors relate this pattern to differences in thermal inertia between ordnance and surrounding materials, but describe the timing as opportunistic rather than statistically validated.
- Detection Pipeline Performance and Acquisition Regimes: 0.488 to 0.664 was the YOLOV11l mAP50-95 increase from 33 m to 15 m acquisition.The two regimes also differed in platform, sensor, site, season, surface material, and source imagery, so 15 m is considered promising for this implementation rather than definitively superior.
- Detection Pipeline Performance and Acquisition Regimes: 0.316 to 0.511 was the RT-DETR-R50 mAP50-95 increase from 33 m to 15 m implementation.The 15 m source images were substantially larger, and fixed 1600 × 1216 pixel input preserved more small-target detail while maintaining compatible feature-map dimensions.
- Discussion: Higher altitudes increase area coverage and reduce collection time, whereas lower altitudes increase the pixels representing small targets.Altitude selection therefore depends on operational context, including the time available to demining or EOD teams.
- Discussion: Representative local imagery should calibrate and independently validate each selected model before operational use.The study presents YOLOV11l and RT-DETR-R50 as alternative end-to-end routes, not as a benchmark proving one architecture is better.
- Operational Decision Support: Detector outputs should prioritize qualified analyst review and follow-on technical survey or EOD assessment, not declare an area free of explosive hazards.Geospatial grouping, false-alarm testing, and operational clearance validation remained future work rather than completed parts of the study.
- Recommendations for Humanitarian Mine Action and Demining: Thermal and RGB imagery should be collected together, with varied substrates, vegetation, target characteristics, and background-only imagery included.The RGB view supplies visible surface context for interpreting thermal anomalies, while dataset diversity addresses factors that can influence thermal contrast and false detections.
Conclusions
The study developed a multi-campaign UAV thermal dataset and evaluated two object detectors for inert ordnance screening. Both models supported candidate detection under controlled conditions, and the resulting recommendations emphasize representative data, operational calibration, and human-led follow-up rather than standalone clearance.
- Dataset Development, Object Detection, and Recommended Use: 5,855 thermal image label pairs, including 918 positive and 4,937 background images, formed the final source inventory across four controlled field campaigns.The study converted raw thermal and RGB imagery into labeled single-class object-detection datasets.
- Dataset Development, Object Detection, and Recommended Use: 0.925 precision, 0.833 recall, 0.877 F1, 0.905 mAP50, and 0.664 mAP50-95 were the strongest YOLOV11l results, obtained with the 15 m dataset.The 15 m dataset produced the strongest reported YOLOV11l performance.
- Dataset Development, Object Detection, and Recommended Use: 0.911 precision, 0.750 recall, 0.823 F1, 0.740 mAP50, and 0.511 mAP50-95 were the strongest RT-DETR-R50 results, also obtained with the 15 m dataset.Both labeled datasets supported candidate detection under controlled conditions, with preserving input detail important for the larger 15 m imagery.
- Dataset Development, Object Detection, and Recommended Use: Model outputs should prioritize analyst review and follow-on technical survey or EOD assessment rather than declare an area free of explosive hazards.The recommendations also call for varied imagery and representative local validation before use.