Digital agronomist: how computer vision and AI save crops from diseases in the early stages

Abstract

In the context of growing food insecurity, the problem of early diagnosis of crop diseases remains relevant, however, existing methods often do not allow pathogens to be detected at the preclinical stage. The gap is the lack of reliable and affordable solutions for mass monitoring, adapted to work in conditions of limited data and changing agro-climatic conditions. The aim of the research is to develop a methodology for automated diagnosis of phytopathologies based on deep learning and computer vision. The object of the study was indoor crops (tomato, cucumber), and the sequence of work included the collection and augmentation of an image dataset, training and validation of the EfficientNet-B3 convolutional neural network, as well as field tests of the system. As a result, a classification accuracy of 98.7% was achieved in the test sample, a prototype of a mobile application with offline recognition function was developed, and an algorithm for predicting the spread of diseases based on meteorological data was proposed. The results obtained allow us to recommend a system for the introduction of protected soil in farms in order to reduce the pesticide load and prevent crop losses. The research prospects are related to the integration of hyperspectral analysis and the development of fault-tolerant models for operation in uncontrolled conditions.