Robustness of YOLO models for object detection in remote sensing images

Remote sensing imagery enables object detection systems to localize and classify targets for critical applications like surveillance and autonomous driving. However, distortions introduced during image acquisition, transmission, or compression degrade the detection performance, posing challenges for real-world applications. This study conducts a comprehensive robustness evaluation of seven state-of-the-art YOLO models, including YOLOv5, YOLOv7, YOLOv8, YOLOv9, YOLOv10, YOLOv11, and the modified YOLOv5 against four common distortions: Additive White Gaussian Noise (AWGN), JPEG and JPEG2000 compressions, and Gaussian blurring. Using the DOTA-v1.0 dataset, we generated 40 distortion test sets (10 levels per distortion type). The obtained results demonstrate that all distortions degrade performance across all evaluated models. YOLOv9 outperforms others YOLO models in terms of mean average precision under different distortions. YOLOv7 and YOLOv10 exhibit the weakest robustness, whereas YOLOv11 shows low resistance to AWGN distortion.