Detection and counting of root-knot nematodes using YOLO models with mosaic augmentation

Abstract

Root-knot nematodes (RKN) are microscopic plant parasites that cause significant economic damage to crops and vegetables. Accurate assessment of RKN populations is required for effective management of this disease. Several “You Only Look Once” (YOLO versions 2 to 7) architectures were investigated for the application of RKN enumeration in microscopic images. YOLOv5-608 model attained Precision score of 0.960, Recall of 0.951, F1-score of 0.990, mAP of 0.972 without mosaic augmentation. Using mosaic dataset, this was increased to Precision of 1.00, Recall of 0.998, F1-score of 0.999, and mAP of 0.995. YOLOv5-608 model showed the highest correlation between the manual and machine counting of RKN: coefficient of determination (R2) of 0.991, root mean square error (RMSE) of 0.313, and coefficient of variation (CV) of 0.251. For free-living nematodes (FLN), this resulted in R2 of 0.994, RMSE of 0.058, and CV of 1.760. YOLOv7-608 achieved the highest correlation between manual and machine counting of overlapped RKN (R2 of 0.970, RMSE of 0.595, and CV of 0.123). In addition, this study explored a new application of mosaic augmentation to analyse microscopic images acquired with different objective lense magnifications. The proposed framework supports the rapid assessment of plant parasitic nematodes necessary to implement nematode control strategies and improve crop management practices.