DIAGNOSTICS OF RUSTS AND SPOTTS OF WHEAT USING COMPUTER VISION METHODS

Abstract

Purpose. To investigate the possibility of diagnosing yellow and brown rust (Puccinia striiformis f. sp. tritici West. Puccinia triticina f. sp. tritici Erikss), yellow spot (Pyrenophora tritici-repentis (Died.) Drechsler) on wheat leaves from images using convolutional neural networks, to compare the most successful and compact neural network architectures. Materials and methods. The material for the research was the images of samples of wheat leaves affected by rust and spots, obtained in the conditions of infectious nurseries of the All-Russian Research Institute of Biological Plant Protection. The total sample size included 5169 images, including brown rust - 227, yellow rust – 1283, yellow spot – 3659. Research methods: data preprocessing methods, training methods for convolutional neural networks. Results. A comparison was made of the four most successful and compact neural network architectures GoogleNet, ResNet-18, SqueezeNet-1.0 and DenseNet-121. On test data – 518 images, all of the above models demonstrated high prediction quality. The best result was shown by DenseNet-121, providing a classification accuracy of over 99 %, with two false positives. Conclusion. The possibility of diagnosing wheat fungal diseases from images using modern methods of computer vision was analyzed. It is shown that under controlled conditions, with a competent organization of the process of collecting and marking data, the problem is successfully solved, and the listed models are the reserve that will automate the stage of phytosanitary monitoring diagnostics.

Keywords

• wheat rust
• wheat speck
• computer vision
• convolutional neural networks