Improving deep learning-based automatic detection and identification of small UAVs using MuSES-generated synthetic imagery

Deep learning has proven effective for detecting and identifying embedded targets in image-based scenes, but achieving high performance requires robust algorithm training. This typically depends on access to a large, diverse dataset for reliable statistical prediction. In the thermal infrared wavebands, obtaining such datasets from measured sources can be particularly challenging, especially when adversarial assets are the focus. To ensure robust training, an accurate image-generation methodology for synthetic but realistic scene simulation can supplement available measured imagery. In this study, synthetic images of several commercial and military unmanned aerial vehicles (UAVs) are generated using physics-based MuSES thermal infrared simulations. The aircraft are simulated with realistic heat sources (e.g., gas engines, batteries and electric motors, aerodynamic heating) under various weather conditions, at multiple times of day, and for many sensor and target positions. This process for generating large and diverse datasets is automated with CoTherm and includes options to incorporate the motion blur of spinning propeller blades and to create composite imagery by inserting synthetic targets into measured background scenes. The performance of YOLO11 and Faster R-CNN detection and recognition algorithms are evaluated using different backgrounds (both sky and terrain) and for various target resolutions to assess effectiveness in detecting and recognizing UAVs in the thermal infrared spectrum. Detection and identification accuracy are examined for real (measured) and synthetic test images for several sensor slant ranges.