Neural networks as a support element of phytosanitary monitoring of fruit crops on the example of apple trees

Abstract

The paper presents the results of developing a convolutional neural network model for detecting and classifying diseases based on images of apple tree leaves and fruits. The study involves transfer learning for the YOLOv10-X model (You Only Look Once, version 10, Extra-large), pre-trained on the public COCO dataset (Common Objects in Context), which includes over 200,000 images and millions of annotated objects. The training dataset was compiled in the Research and Production Department of the Federal Horticultural Center for Breeding, Agrotechnology and Nursery (Russia). Artificial augmentation of the training dataset by rotating images, adding noise, and changing tints and shades increased the dataset to 2200 images. The Precision and Recall metrics, as well as the mean Average Precision (mAP) metric, were used to evaluate the performance of the model. The study demonstrated that the model effectively recognizes leaf lesions caused by scab, powdery mildew, rust, and various types of spots, achieving a mean Average Precision (mAP) of 0.6. The “spot” class appeared to be the most difficult to recognize (mAP50=0.411; Recall=0.324), while the “rust” class revealed the least difficulty (mAP=0.868; Recall=0.803). The study contributed to optimizing the model parameters, including the confidence threshold (0.48), the learning rate (0.01), the number of epochs (313) and the batchsize (8). Testing of a robotic platform equipped with RGB cameras indicated that automatic data collection at high frequency enables effective real-time monitoring of lesion dynamics.