Classification of crop infestation with field bindweed using ResNet neural network architectures

Abstract

The development of effective methods for automatic weed monitoring is one of the key tasks for improving the sustainability, productivity, and profitability of modern agriculture. The introduction of such technologies underlies the concept of precision farming, which aims at optimizing resources and minimizing anthropogenic impact on agroecosystems. The key role in solving this problem is played by artificial intelligence technologies, in particular Deep Learning methods. The study is devoted to a comparative analysis of the applicability of convolutional neural networks (CNNs) with ResNet-18, ResNet-34, and ResNet-50 architectures for solving the problem of classifying crops of five agricultural species (wheat, barley, buckwheat, rapeseed, and flax) based on the sign of infestation with field bindweed. All the tested architectures demonstrated high efficiency, achieving accuracy of more than 91% and an F1-score in the range of 0.85–0.90, which indicates a good balance between the accuracy and completeness of predictions. It was found that the best results were achieved by not the deepest model – ResNet-18, which, despite having lower computational complexity, outperformed the more powerful counterparts ResNet-34 and ResNet-50 in stability and overall prediction accuracy. This fact