UAV-based Terrestrial Oil Spill Dataset

UTOSD: UAV-based Terrestrial Oil Spill Dataset UTOSD is a UAV-based terrestrial oil spill semantic segmentation dataset designed for developing and evaluating deep learning methods for land oil spill detection. The dataset contains 665 high-resolution RGB images captured using DJI Matrice-series UAVs over oilfields and their surrounding environments. The images were collected under diverse environmental conditions, including multiple flight altitudes, illumination conditions, and acquisition times. Representative terrestrial scenes include grasslands, bare soil, snow-covered areas, and oilfield facilities, providing a wide variety of challenging backgrounds for semantic segmentation. Pixel-level annotations were manually created using LabelMe and subsequently converted into Pascal VOC binary segmentation masks. In the annotation masks, a pixel value of 0 represents the background, while 1 represents the oil spill class. The dataset is organized into three corresponding folders: 1. JPEGImages/: Original RGB images. 2. labels/: Original LabelMe annotation files (.json). 3. SegmentationClass/: Pascal VOC binary segmentation masks (.png). Each image, JSON annotation, and binary mask share the same filename to ensure one-to-one correspondence. UTOSD is intended for research on UAV remote sensing, semantic segmentation, lightweight vision models, and intelligent monitoring of terrestrial oil spills.