Deep Object Detection of Crop Weeds: Performance of YOLOv7 on a Real Case Dataset from UAV Images

Weeds are a crucial threat to agriculture, and in order preserve crop productivity, spreading agrochemicals is common practice with potential negative impact on the environment. Methods that can support intelligent application needed. Therefore, identification mapping critical step performing site-specific weed management. Unmanned aerial vehicle (UAV) data streams considered best for detection due high resolution flexibility of acquisition spatial explicit dimensions imagery. However, existence unstructured conditions biological variation weeds, it remains difficult challenge generate accurate recognition models. Two barriers tackling this related (1) lack case-specific, large, comprehensive UAV image datasets interest, (2) defining most appropriate computer vision (CV) models assess operationality approaches real case conditions. Deep Learning (DL) algorithms, appropriately trained deal complexity provide valid alternative solutions respect standard CV an model. In framework, paper first introduces new dataset named Chicory Plant (CP) then tests state-of-the-art DL algorithms object detection. A total 12,113 bounding box annotations were generated identify targets (Mercurialis annua) from more than 3000 RGB images chicory plantations, collected using system at various stages growth. was conducted by testing recent You Only Look Once version 7 (YOLOv7) both CP publicly available (Lincoln beet (LB)), which previous YOLO used map weeds crops. The YOLOv7 results obtained encouraging, outperforming other variants producing value metrics 56.6%, 62.1%, 61.3% [email protected] scores, recall, precision, respectively. Furthermore, model applied LB surpassed existing published increasing scores 51% 61%, 67.5% 74.1%, 34.6% 48% mAP, mAP sugar beets, This study illustrates but remarks fundamental needs large-scale, annotated develop evaluate real-case field circumstances.