Comparative Analysis of Military Detection Using Drone Imagery Across Multiple Visual Spectrums

Technical Summary: Comparative Analysis of Military Detection Using Drone Imagery Across Multiple Visual Spectrums

Problem Statement
Modern military operations increasingly rely on drones for surveillance, reconnaissance, and targeted strikes. While drone technology has advanced, the reliability of object detection systems remains a critical challenge when operating in diverse and hostile environmental conditions. Existing models often perform well in controlled settings but struggle with accuracy in real-world scenarios characterized by low-light, thermal variations, fog, or motion distortion. Current research lacks extensive evaluation of detection robustness across varying spectral conditions (thermal, night vision, grayscale) and obscured environments, creating a gap in developing adaptive systems for dynamic military operations.

Methodology
To address these challenges, the authors utilized the KIIT-MiTA dataset, a collection of 1,700 high-quality drone-captured images containing over 4,100 annotated labels across seven military classes: Artillery, Missile, Radar, Multiple Rocket Launcher, Soldier, Tank, and Vehicle.

The core methodology involved transforming this standard dataset into four distinct visual modalities to simulate real-world operational constraints:

1. Thermal Vision: Images were converted to grayscale and mapped to a pseudo-thermal color space using the OpenCV COLORMAP_INFERNO to simulate heat signatures.
2. Night Vision: Images were converted to grayscale, enhanced for brightness and contrast, and overlaid with a green tint to mimic monochromatic image intensifiers used in low-light surveillance.
3. Gray Scale: Standard conversion to a single intensity channel to remove color distractions and focus on spatial features.
4. ObscuraVision: A specialized augmentation simulating real-world visibility challenges (light fog, motion blur, contrast irregularities) using a weighted combination of motion blur, fog coefficients, and contrast/brightness adjustments.

The YOLOv11-small model was selected for its lightweight architecture and efficiency in real-time detection. The model was trained on a cloud environment (NVIDIA T4 GPU) for 100 iterations with a batch size of 16 and 640×640 pixel input. Training employed image augmentation, a learning rate scheduler, and optimization of classification, objectness, and bounding box regression losses.

Key Contributions

• Dataset Extension: The creation of four distinct dataset versions (Gray Scale, Thermal, Night Vision, ObscuraVision) derived from the KIIT-MiTA dataset to simulate specific environmental challenges.
• Comparative Evaluation: A systematic assessment of the YOLOv11-small model's performance across these four modalities, focusing on how environmental factors impact detection accuracy.
• Performance Benchmarking: The establishment of baseline metrics (mAP@50, mAP@50-95, Precision, Recall, F1-Score) for military object detection under simulated thermal, nocturnal, and obscured conditions.

Results
The study evaluated the models using standard detection metrics, with a primary focus on Mean Average Precision (mAP). The results indicated significant variations in performance based on the visual modality:

• Night Vision achieved the highest detection accuracy, with an mAP@50 of 0.701 and mAP@50-95 of 0.484. However, it incurred the highest total processing time (10.1 ms).
• ObscuraVision followed closely with an mAP@50 of 0.694 and mAP@50-95 of 0.467, offering a strong balance between accuracy and inference speed (9.4 ms).
• Thermal Vision recorded an mAP@50 of 0.680 and mAP@50-95 of 0.466, performing competitively despite the challenges of thermal imagery.
• Gray Scale demonstrated the fastest processing time (8.6 ms) but yielded the lowest detection accuracy, with an mAP@50 of 0.603 and mAP@50-95 of 0.374.

While the Gray Scale model achieved the highest F1 Score (0.700) and Thermal Vision the highest Precision (0.752), the study prioritized mAP metrics for model ranking. A confusion matrix analysis on the Gray Scale dataset further illustrated the model's ability to distinguish between the seven military classes.

Significance and Claims
The paper claims that this research advances the robustness of drone-based detection systems by validating the YOLOv11-small model's adaptability across diverse environmental conditions. The findings suggest that Night Vision is the most suitable modality when detection accuracy is the paramount concern, while ObscuraVision and Thermal Vision offer favorable trade-offs between performance and speed. Conversely, Gray Scale is identified as appropriate for latency-sensitive applications where lower accuracy is acceptable.

The authors conclude that the study provides valuable insights for improving defensive and offensive military operations, ensuring continuous effectiveness in dynamic environments. They note that future work could explore multi-modal fusion, temporal data integration for motion-aware tracking, and optimization for edge devices, but the current study focuses strictly on the comparative evaluation of static image modalities.