Abstract
Flea Beetle Damage (FBD), Deep Learning, You Only Look Once(YOLO), Object Detection, Automatic quantification
Abstract. Accurate estimation of pest damage in agricultural fields is crucial for implementing timely and effective management strategies. This study focuses on developing an integrated, scalable, cost-effective framework for in-field canola flea beetle damage estimation using high-resolution multimodal imaging and advanced deep-learning techniques. Ground-level smartphone high-resolution images were analyzed to segment and quantify damage to canola leaves using the YOLO algorithm. YOLOv8 and YOLOv10 were applied and utilized to segment the background, leaf area, and flea beetle-damaged region of interest. The analysis can be further expanded by utilizing drone imagery to compute vegetation indices and assess damage at a larger scale. The segmented damage areas were correlated with ground-truth observations for validation and scaling. 70% of the high-resolution images of plants with beetles and beetle damage are used in training and 25% of the images are used in testing, and the rest of the images are used for validation purposes. The damaged leaf area was delineated and computed on a pixel-by-pixel basis to identify and categorize three main components: flea beetle damage, healthy leaf tissue, and the occurrence of crucifer flea beetles. This segmentation approach provides a more accurate and thorough assessment of the damage in comparison to conventional scouting techniques. The YOLOv10 model's test results showcased remarkable performance, attaining 99% overall accuracy, with 98% accuracy in detecting beetle damage, 100% accuracy in leaf detection, and 100% accuracy in recognizing the presence of crucifer flea beetles.