Real-Time Detection of Shot-Hole Disease in Cherry Fruit Using Deep Learning Techniques via Smartphone

Abstract

Nowadays, pesticides are generally used to control diseases and pests. However, many farmers often do not fully understand what diseases and pests are and the extent of their effects. For this reason, the optimal use time of pesticides may be missed, or excessive amounts of pesticides may be used. For this reason, early detection and identification of the disease and pest should be made. One of the methods that allows early detection is deep learning. In this study, deep learning methods were used to detect shot-hole disease, which causes damage to the fruit part of the cherry tree, one of the Prunus species, in real time via a smartphone. To achieve this determination, studies were first carried out on object recognition algorithms in three different methodologies. These models are YOLOv8s, DETR Transformer and RTMDet MMDetection. In the training and test results performed on the created hybrid dataset, it was seen that the most successful algorithm was YOLOv8s. For the YOLOv8s algorithm, mAP₅₀, mAP_50-95, precision and recall performance metrics were found to be 92.7%, 58.9%, 86.7% and 90.2%, respectively. Since YOLOv8s showed the highest successful performance, this algorithm was used in the study for real-time detection. In the real-time experiment, it was determined that it correctly detected 115 of 119 images on the test dataset with an F1 score value of over 80%. As the output of the study, a QR (Quick Response) code was created in the study so that real-time detection can be attempted with a smartphone.