Deep learning classification of INSV-associated weeds in Monterey county using a curated RGB image dataset

Area of Science:

• Agricultural Science
• Computer Science
• Plant Pathology

Background:

• Weeds cause significant crop losses by competing for resources and spreading diseases.
• Sonchus oleraceus and Malva parviflora are linked to $150 million in losses in Monterey County due to Impatiens Necrotic Spot Virus (INSV).
• Precision agriculture requires advanced tools for early weed detection and disease management.

Purpose of the Study:

• To develop a region-specific image dataset for INSV-associated weeds in Monterey County.
• To evaluate the performance of deep learning models for classifying visually similar weeds.
• To establish a foundation for real-time weed detection systems in high-value crop production.

Main Methods:

• Created a high-resolution image dataset of Sonchus oleraceus and Malva parviflora under controlled greenhouse conditions.
• Compared ResNet-50, ResNet-101, and DenseNet-121 convolutional neural networks for weed classification.
• Utilized data augmentation and ten stratified data splits for robust model training and validation.

Main Results:

• ResNet-101 achieved the highest median classification accuracy (91%) and Cohen's Kappa (0.87).
• DenseNet-121 showed superior F1-score and Area Under the Curve (AUC) values (>0.99).
• Dataset augmentation significantly improved model generalization and classification performance.

Conclusions:

• Deep learning models are effective for accurate weed identification, even with visually similar species.
• The developed dataset addresses a critical gap for California's high-value crop systems.
• This research supports the development of sustainable, targeted weed control strategies in precision agriculture.