AUTOMATIC DETECTION AND IDENTIFICATION OF WHEAT DISEASES USING DEEP LEARNING AND REAL-TIME DRONES

Abstract

The study was carried out to develop a method for early detection of wheat diseases using a modified version of the YOLOv8n model, aimed at achieving an optimal ratio of speed and accuracy of detection in real time. To increase the accuracy of the regression block for recognizing diseases such as brown rust, yellow rust, mold and septoria, the GIoU bounding box regression loss function has been introduced. A simplified YOLOv8n network structure is proposed, adapted for use as a backbone network to facilitate broadcasting to mobile terminals. The use of pretraining methods that combine blended and transfer learning helps improve the model’s generalization ability. For the analysis, we used data collected during four field experiments in Ufa and Karmaskalinsky districts of the Republic of Bashkortostan. In total, more than 10.000 images were collected during the field experiment, of which 1.890 images of wheat leaves were selected for model training. Data processing included statistical analysis of images obtained under various lighting conditions. Recognition and evaluation of model efficiency were carried out using F1 and AP indicators. The F1-score when testing the model on images taken against a background of sufficient lighting and without covering by leaves was 54%, and the AP-score was 51.2%, with an average IOU value of 50%. The accuracy of wheat disease identification in images of the training data set reached 60%, and of the test set – 70%. The developed YOLOv8n model for detecting wheat diseases in field conditions demonstrates the ability to identify diseases in real time with an accuracy of 67.53%, which significantly exceeds the value of this indicator in other existing models. In particular, the accuracy of YOLOv8n is 3 times higher than that of the YOLOv4 model, indicating significant progress in the field of automatic plant disease detection.