Restyling Synthetic Multispectral Imagery with CycleGANs for UAS-Based Landmine Detection

Abstract

In 2024, landmines caused over 6,000 recorded casualties, many of them children. Prior work has shown Unmanned Aerial Systems paired with multispectral imagery and computer vision to be a viable detection method, but such systems require large amounts of labeled minefield imagery that are costly to collect. We investigated whether a generative adversarial network (GAN) can close the visual gap between computer-rendered synthetic multispectral imagery and real aerial RGB-LWIR captures well enough to substitute for real training data. We trained a CycleGAN, conditioned on RGB-LWIR fusion level, season, time of day, and altitude, to restyle Blender-rendered frames and, as a control, real imagery, toward 7,517 real RGB-LWIR images from the AMLID dataset. Refinement reduced the mean Earth Mover's Distance between generated and real intensity histograms to 5.74, surpassing unrefined renders (7.09). We trained YOLOv11 and RF-DETR on six sets replacing 30%, 50%, or 70% of real imagery with GAN-refined renders or GAN-refined real images, testing on 1,320 real images across 264 stratified conditions. Replacing half the real set with GAN-refined real imagery cost only 1.7-2.6 mAP@50 points, but GAN-refined renders cost 17.3-19.9 points, despite comparable histogram realism. Accuracy depended more on survey conditions than training mix, falling 29 points from 5m-20m altitude and declining as LWIR content replaced RGB. GAN refinement lets real imagery be halved at almost no cost, but applying it to synthetic scenes leaves an 18-point gap, meaning that realistic rendering, not just restyling, remains the bottleneck.