Deep Convolutional Neural Network Model for Weed Identification in Oil Flax Crops

Abstract

Traditional methods of monitoring weeds in crops are not suitable for integration with modern "smart" agricultural machinery. Automating the process of accurately identifying the species composition of weeds in each field will be an important step for the development of a plant protection system, contributing to higher yields. Deep learning (DL) models, which have become widespread in recent years, are successfully helping to solve this complex agricultural problem. In this study, we built a classifier based on the ResNet-18 deep learning model, which is able to detect weeds with the corresponding weediness gradations in photographs from field plots with oil flax (Linum usitatissimum L.). There are 4 types of weeds in crops of oil flax with different intensity - field bindweed (Convolvulus arvensis), white goosefoot (Chenopodium album), leafy spurge (Euphorbia virgata) and wild buckwheat (Fallopia convolvulus). The task of the classifier is to recognize these weeds in the photograph and determine one of the two gradations of weediness of the plot - the number of weeds exceeds the economic injury level (EIL) or does not exceed. The models were trained with different epoch values (10, 20, 30), the accuracy of which ranged from 72.5 to 93.3%.