Towards AI-based airborne field emergence assessment of grain crops: an experience report

This paper reports on our experiences using unmanned aerial vehicles (UAVs) to collect RGB images for automated assessment of spring barley crops, thereby highlighting challenges and proposing potential solutions. Experiments involved capturing images with UAVs on barley fields at early growth stages BBCH 1 and 2. Challenges we encountered comprise narrow flight windows due to weather, unforeseen UAV technical deficiencies, and problematic lighting causing overexposure and shadows, making plants unrecognizable. Usually, the whole plant is marked in the labeling process. Here, an alternative approach was used. The YOLO object detector was trained using the seed emergence point (SEP), which was annotated using LabelMe and circles instead of bounding boxes. The annotated data was processed with a fine-tuned YOLOv9e object detector, resulting in poor mAP50 of 28 %. This is attributed to the low quality of the collected image data, which is limited due to the technical challenges we have been facing. With this contribution, we intend to share our experiences in the hope of preventing other groups from running into similar issues.