Digital monitoring of crops in grain ecosystems

Abstract

In the conditions of rapid global population growth, resource depletion, and increasing demand for grains, an efficient agricultural management system becomes a crucial element for ensuring food security in Russia and worldwide. The foundation of such management is an intelligent grain production monitoring system, where diagnosing grain crop diseases serves as a critically significant subsystem. This article presents an approach based on the utilization of neural networks, specifically the U-Net architecture for semantic segmentation, adapted for the analysis and detection of helminthosporium through images of maize leaves. Quality evaluation of segmentation employs metrics like Intersection over Union (IoU) and Dice coefficient, computed from a held-out dataset, ensuring an objective assessment of results. The research demonstrates high accuracy and similarity between the model's predictions and expert annotations, while also showcasing the convergence of loss function during neural network training. A notable advantage of the proposed approach lies in the lightweight nature of the suggested architecture and the ability to utilize trained models as cores for decision support systems, including on local devices without network connectivity.