Vehicle Detection from High-Resolution Satellite Images Using YOLOv8 for Traffic Monitoring

Abstract

Recent availability of high-resolution space-borne satellite imagery has opened new opportunities for large-area vehicle monitoring, traffic analysis, urban planning and intelligent city management. Automated vehicle detection in satellite imagery still remains quite difficult owing to different scales of vehicles, the prevalence of complex backgrounds occlusion shadows, lighting variations, and factors leading to huge scale variance. This paper proposes a deep-learning driven vehicle-detection structure based on YOLOv8 object detection structures to enhance the vehicle detection performance, in particular for small vehicle detection in high-resolution satellite images. This setup adopts multiple image resizing and adaptive image tiling mechanisms during the detection cycle to maintain spatial relation information of detected boxes, this way mitigating the small object detection problem. It further employs data augmentations and multi-scale feature extraction mechanisms to preserve detail information for a complete detection process. Considering the use of publicly available aerial imagery datasets (e.g. VEDAI COWC etc.), performance is evaluated using precision recall mAP, F1-score and inference speed. The motivation of this research is based on the results reported among 16 selected vehicle-detection studies over 20192025, which consistently identified that tiny objects, multiple vehicles clusters, occlusion and large imagery size remained major problems in detection tasks. The publicly available EAGLE dataset can provide >200,000 vehicles as training samples. The proposed YOLOv8 setup aims to enhance small vehicle detection accuracy, generalization and robustness, without or with marginal compromise to efficiency. This will help achieve the requirements for large-area vehicle detection, transport research, mobility analysis, smart city planning and other remote sensing applications. This study offers a quantitative structure of vehicle detection evaluation metrics for YOLOv8 model suitability assessment.