Artificial Intelligence Based Tools for Automated Genus-Level Identification of Plant-Parasitic Nematodes

Early and accurate identification and quantification of plant-parasitic nematodes (PPN) is crucial for their effective control. Although valuable, the current techniques for identifying PPN, such as morphology and molecular marker-based methods, can be time and resource-intensive. This study aims to develop and validate cutting-edge computer vision tools for automated, accurate, and reproducible PPN detection. To achieve this goal, we captured microscopic images of the three economically-important PPN genera associated with potato crop: root lesion (RLN; <i>Pratylenchus</i> spp.), root-knot (RKN; <i>Meloidogyne</i> spp.), and stubby root (SRN; <i>Paratrichodorus</i> and <i>Trichodorus</i> spp.), additional plant-parasitic nematodes (PPN-OTHERS) and non-parasitic (NON-PARASITIC) nematodes, for a total of five groups. The captured images (total instances = 8,654) were preprocessed, annotated, and randomly split into three datasets: 75% for training, 15% for validation, and 10% for testing. An object segmentation algorithm, YOLOv11-seg, which predicts each pixel in the image, was trained and evaluated on previously unseen images. The model achieved high accuracy in validation (92.4%) and test: (88.6%) datasets with strong performance for key PPN genera (RKN, RLN, SRN; F1-scores >0.92; AUC >0.93 in the test set). While the NON-PARASITIC showed strong performance (F1-score > 0.846 and AUC >0.91), the PPN-OTHERS group performed poorly (test accuracy: 43.9%), frequently misclassified as RLN and NON-PARASITIC nematodes. The results highlight the potential of artificial intelligence-based tools in identifying PPN, paving the way for the long-term goal of developing automated detection and quantification systems for plant pathogens.