Can computers effectively detect crop pests? Evaluating a deep-learning model for watermelon pest detection on sticky cards

Cucumber beetles are significant pests in specialty crops such as watermelon, yet no dedicated computer vision models currently exist for their detection. This study evaluates the performance of deep learning models, YOLO and RT-DETR, in identifying cucumber beetles (spotted and striped) on yellow sticky traps, a common monitoring method in agricultural pest management. A dataset of 375 manually annotated images were used and expanded through patching and augmentation, incorporating variations in lighting, contrast, blur, and noise to simulate real-world challenges. Using a five-fold cross-validation framework, models were trained and optimized via grid search to assess precision, recall, and F1 score across confidence thresholds. YOLO demonstrated faster inference and achieved perfect precision (100%) at an 84.3% confidence threshold, while RT-DETR showed superior overall performance, reaching an F1 score of 97% at 83.3% confidence. Both models successfully distinguished cucumber beetles from visually similar species. These findings demonstrate the potential of computer vision systems in automating pest detection with high accuracy and offer a scalable solution for integrating AI into field-based agricultural monitoring. Ongoing efforts include deployment testing under diverse field conditions to evaluate model generalizability and robustness.