Abstract
Weeds, defined by their ecological and economic impact rather than taxonomy, present a major challenge to agriculture by competing with crops for limited resources and serving as vectors for disease. In Monterey County, California, one of the most productive farming regions in the United States, Sonchus oleraceus (annual sowthistle) and Malva parviflora (little mallow) have been linked to over $150 million in crop losses due to their role in spreading Impatiens Necrotic Spot Virus (INSV). As precision agriculture becomes more important in high-value production systems, deep learning and image-based classification offer promising tools for early weed detection and disease prevention. To address the absence of region-specific image datasets, this study presents the first curated, high-resolution image collection of INSV-associated weeds from Monterey County, captured under greenhouse conditions designed to mimic field variability. This dataset fills a documented gap in existing global repositories such as PlantCLEF and DeepWeeds, which lack representation of California's high-value crop systems. This study compares three convolutional neural networks-ResNet-50, ResNet-101, and DenseNet-121-for classifying these visually similar weeds under controlled conditions that approximate real field environments. RGB images were augmented to improve model robustness, and training was conducted across ten independent stratified data splits.Among the tested architectures, ResNet-101 achieved the highest median classification accuracy (91%) and Cohen's Kappa (0.87), while DenseNet-121 demonstrated the strongest F1-score and AUC values exceeding 0.99. These results confirm that dataset augmentation substantially enhanced model generalization. The results demonstrate that deep learning can support accurate and reliable weed identification, paving the way for real-time detection systems and more targeted, sustainable weed control practices in precision agriculture.