Deep learning for automated weed monitoring in grain crops

Crop yield reduction under the influence of biotic stressors such as weeds remains a pressing issue in agroecosystems. Traditional methods of weed infestation monitoring based on visual assessment are laborintensive and subjective. In this paper, an approach to automated weed identification in spring wheat and barley crops using ResNet convolutional neural networks (CNNs) is proposed. The objective of the study was to develop classifiers based on ResNet-18, ResNet-34, and ResNet-50 architectures to detect 16 weed species and determine the need for herbicide treatment. The dataset included 138 images with a resolution of 1340 × 1790 pixels obtained from a mobile camera and phytosanitary monitoring data from 66 survey plots (0.25 m²). To compensate for the small amount of data, augmentation was used (aug_transforms library, PyTorch) with the following operations: random rotations, scaling, brightness and contrast correction. This allowed us to expand the sample by at least 5 images per class. The images were preprocessed: scaling to 512 × 512 pixels with subsequent compression to 224 × 224 for compatibility with ResNet. The models were trained for 100 epochs with the Adam optimizer (batch size = 16), the quality metrics were accuracy_multi (the proportion of correct classifications) and F1-score. All models achieved high accuracy (accuracy_multi > 95%), but the F1-score ranged from 0.60 to 0.74, which reflects the complexity of multi-class classification. ResNet-18 demonstrated the highest F1 scores (0.74). Confusion matrix analysis revealed problematic classes: Convolvulusarvensis (6 errors), Sinapisarvensis (4 errors), which is due to insufficient data representativeness. Classes 4, 7–10, and 14 were recognized with the highest accuracy. Confusion Matrix also revealed ambiguous effectiveness of deep ResNet architectures.