Optimizing YOLOv12-L for Imbalanced Satellite Vehicle Detection via Physics-Informed Data Synthesis and Adaptive Class-Aware Loss Weighting

Vehicle detection in satellite imagery underpins applications in traffic monitoring, urban planning, and disaster response, yet two obstacles routinely degrade detector performance: the small pixel footprint of individual vehicles and severe imbalance between common and rare vehicle categories. This study presents a vehicle detection framework built on the YOLOv12-L architecture and evaluated on the Vehicles in the Middle East (VME) dataset, in which cars outnumber buses and trucks by a wide margin. The framework combines a physics-informed synthetic data generator, which places procedurally rendered bus and truck silhouettes onto realistic backgrounds and enhances them with adaptive contrast and sensor-noise simulation, with a class-aware loss configuration that increases the classification loss weight and applies label smoothing for minority categories. Training on 4,666 images (2,828 original tiles plus 1,838 synthetic tiles contributing 2,476 bus and 2,076 truck instances) for 100 epochs produced a validation mean Average Precision at an IoU threshold of 0.5 (mAP50) of 0.687 and mAP50:95 of 0.391, with per-class mAP50:95 values of 0.476 for cars, 0.374 for buses, and 0.323 for trucks. On a held-out test split, mAP50 reached 0.634 and mAP50:95 reached 0.361. Increasing the inference resolution from 640 to 736 pixels raised validation mAP50:95 to 0.398 without retraining. Test-Time Augmentation at the 1.17× scale further improves overall mAP50 to 0.693. These results confirm that targeted data synthesis combined with class-aware loss weighting yields consistent gains for underrepresented vehicle categories in geospatial imagery.