Semantic segmentation of rusts and spots of wheat

Abstract

The paper explores the possibility of semantic segmentation of the yellow rust and wheat blotch classification using the U-Net convolutional neural network architecture. Based on an own dataset of 268 images, collected in natural conditions and in infectious nurseries of the Federal Research Center for Biological Plant Protection (VNII BZR), it is shown that the U-Net architecture with ResNet decoders is able to qualitatively detect, classify and localize rust and spotting even in cases where diseases are present on the plant at the same time. For individual classes of diseases, the main metrics (accuracy, micro-/macro precision, recall, and F1) range from 0.92 to 0.96. This indicates the possibility of recognizing even a few diseases on a leaf with an accuracy that is not inferior to that of a plant pathology expert. The IoU and Dice segmentation metrics are 0.71 and 0.88, respectively, which indicates a fairly high quality of pixel-by-pixel segmentation and is confirmed by visual analysis. The architecture of the neural network used in this case is quite lightweight, which makes it possible to use it on mobile devices without connecting to the network.