Multi-Background UAV Spraying Behavior Recognition Dataset for Precision Agriculture

This dataset contains 9,548 high-quality annotated images for recognizing agricultural UAV spraying behaviors under diverse field conditions. It is constructed to address the lack of domain-specific datasets for operational state recognition in precision agriculture, where existing UAV datasets mainly focus on target localization in urban security or airspace management scenarios. Images were collected from 71 online videos covering real agricultural spraying operations and retain heterogeneous resolutions and shooting conditions to reflect real-world monitoring environments.The dataset covers six representative agricultural background types, including green cropland, bare farmland, orchards, woodlands, mountainous terrain, and sky scenes, and exhibits challenging characteristics such as small targets, large scale variation, motion blur and complex, dynamic backgrounds. A dual-state annotation strategy is adopted, where each UAV instance is labeled as either \u201cspraying\u201d or \u201cflying without spraying\u201d, with precise bounding boxes and rich metadata in both JSON and TXT formats. In total, 5,687 instances are annotated as spraying and 4,027 as non-spraying flight, providing an almost balanced distribution of operational states.Benchmark experiments using mainstream object detection algorithms (Faster R-CNN, SSD, YOLOv5 and YOLOv8) demonstrate that the dataset can effectively reveal performance differences under agricultural scene challenges. The lightweight YOLOv5n model achieves a precision of 97.86 and mAP50 of 98.30 on this dataset, confirming its suitability for training and evaluating spraying-state recognition models in precision agriculture monitoring and supervision applications.