大豆圃場における空撮画像を用いた機械学習による微小アサガオ検出の自動化

―植生指数や形状などに基づいたアノテーション―

抄録

In soybean fields, morning glories have a negative impact on soybean growth and yield. In recent years, research has been conducted to use machine learning to recognize weeds using aerial images of fields taken by drones, but it is difficult to collect data on small morning glories that have just germinated. In addition, the data must be sorted and color-coded one by one, making the annotation process time-consuming. Therefore, in this study, a drone was used to take images of soybean fields in the early stages of cultivation and collect image data of morning glories using color information of crops and weeds, and the annotation process was automated. For the automated annotation, we used two thresholds that are likely to respond to morning glories, using color information and shape differences among soybeans, morning glories, and grass weeds. Training on the automatically annotated dataset successfully detected small morning glories.