DeVCL: An end-to-end degradation-aware framework for vehicle counting and localization in satellite imagery

Vehicle counting and localization using high-resolution satellite imagery have recently demonstrated substantial value in urban management and public services. However, satellite images routinely suffer from degradation issues, including blurring, insufficient resolution, noise, uneven illumination, and occlusion, due to sensor limitations, weather conditions, compression artifacts, and other environmental factors. These quality issues severely degrade the stability and accuracy of traditional vehicle counting and localization methods. To address this challenge, we propose Degradation-aware Vehicle Counting and Localization (DeVCL), a novel end-to-end point regression framework explicitly designed for degraded satellite imagery, which adaptively recognizes image degradation conditions and directly predicts vehicle positions. Specifically, DeVCL uses a self-supervised degradation representation jointly with image quality assessment to guide a degradation-aware feature modulation module, enhancing feature representations for low-quality inputs. We also introduce a feature-level adversarial mechanism without paired supervision to strengthen feature robustness. In addition, a density-sensitive feature refinement module is proposed to address matching ambiguities caused by densely packed and arranged vehicles, thus improving localization performance. We evaluated DeVCL using two synthetic degraded datasets built on FAIR1M_V and ITCVD, along with a newly collected dataset named SatPark that features high vehicle density and includes multiple naturally occurring degradations. Experimental results indicate that DeVCL consistently outperforms existing methods, particularly in SatPark, demonstrating strong generalization and practical adaptability. • DeVCL is an end-to-end vehicle localization framework for degraded satellite images. • Construct a public dense satellite parking dataset with natural degradations. • Adaptively identifies LQ images and performs degradation-aware modulation. • Unpaired adversarial alignment aligns modulated features to HQ features. • Density-sensitive refinement improves localization in dense vehicle scenes.