An Analysis of the Impact of Physically-Based Fog Synthesis on YOLOv10s-Based Satellite Object Detection Performance

In satellite-based object detection, fog and low illumination significantly degrade recognition performance. This study investigates the effect of a physically-based fog synthesis(PBFS) method on the detection performance of YOLOv10s. Three datasets were used for training: original images, brightness-enhanced images using OpenCV, and fog-synthesized images generated through a depth-based atmospheric scattering model. Evaluation was conducted using APs, APm, and APl, with FPS, latency, and VRAM analyzed as auxiliary computational metrics. The results show that the model trained with physically synthesized fog images achieved the best performance among the compared conditions, demonstrating improved robustness under foggy and low-illumination conditions. Compared to conventional brightness-based augmentation, the PBFS method provided more realistic visibility degradation and enhanced adaptability to environmental variability. These findings highlight the potential of physics-based image synthesis for reliable satellite target detection in military and environmental monitoring applications.