Real-Time Oil Spill Detection for UAV Images in Varied Terrains in Iraq Based on Deep Learning

Abstract

Oil spills pose serious threats to both the environment and public health, demanding fast and accurate detection in all settings. Traditional detection methods are often slow, expensive, and ineffective across diverse terrains. This highlights the urgent need for an efficient, automated solution capable of reliably identifying oil contamination in real time. In this paper, we developed an aerial system using YOLOv10 detector to locate oil contamination on and off the Tigris River banks near Tikrit, Iraq, in real-time. We collected and annotated a custom dataset comprising 700 UAV images, consisting of 470 water-based oil spill images and 230 land-based oil spill images. The images underwent square cropping and normalization, followed by a comprehensive augmentation pipeline that included mosaic mixing and class-aware sample selection. Fine-tuning COCO-pre-trained YOLOv10 weights using freeze-then-unfreeze strategies yielded water-only mAP₅₀ results of 69.1 % and land-only 63.1 %. For the combined test set, the detector achieved mAP₅₀ 79.7 %. These findings confirm the lightweight UAV platform and the single-stage detector are suitable for extensive real-time environmental monitoring in deprived regions with limited resources.