Abstract
<jats:p>This article presents annotated Soybean_2021_Almalybak dataset designed for automatic analysis of agricultural areas based on aerial photographs obtained using unmanned aerial vehicles (UAVs). The dataset includes images of soybean fields covering various stages of plant growth and containing manual labelling of 12 object classes, including cultivated and weed species. The work emphasises the importance of field data reflecting the realistic state of agricultural landscapes for training robust computer vision models. YOLOv8x, a modern architecture for simultaneous object localisation and classification, was used as the base model. The model was trained on a labelled sample of 300 images without augmentation, which allowed for a benchmark accuracy assessment. The experimental results demonstrated the high efficiency of the model: the F1-score for the Glycine max (soybean) class was 0.933, and the metric mAP@0.5 was 0.72. Despite the limited amount of data and partial annotation, the model showed resistance to variability in field conditions and class imbalance.</jats:p>