Optimizing data collection requirements for machine learning models in wild blueberry automation through the application of DALL-E 2

Abstract

This research developed a workflow to assess the viability of AI-generated imagery in training machine learning models for detecting ripe wild blueberries (Vaccinium angustifolium Ait.), hair fescue weeds (Festuca filiformis Pourr.), and red leaf disease (Exobasidium vaccinii). Ground truth images were collected and augmented with AI-generated variations using DALL-E 2 to expand the dataset. Models were trained on three datasets: ground truth, generated, and a combination (40% generated images). Evaluation metrics included precision, recall, mAP₅₀, and mAP_50–95, analyzed using ANOVA multiple mean comparisons and Tukey's HSD test (α = 0.05). For ripe wild blueberries, combination models achieved the highest performance across all metrics (mAP₅₀: 0.834), significantly outperforming the ground truth model (mAP₅₀: 0.806) in terms of mAP_50–95 (0.478 compared to 0.424). For hair fescue weeds, the combination dataset outperformed others with the highest mAP₅₀ (0.983), closely followed by the ground truth dataset (mAP₅₀: 0.969). In detecting red leaf disease, the combination dataset showed the best performance (mAP₅₀: 0.848 ± 0.140, mAP_50–95: 0.607 ± 0.219), compared to the ground truth (mAP₅₀: 0.615 ± 0.092, mAP_50–95: 0.417 ± 0.045) and generated datasets (mAP₅₀: 0.245 ± 0.088, mAP_50–95: 0.144 ± 0.059). Models trained solely on generated images showed significantly lower performance across all categories except the precision metric for red leaf, where performance was comparable to ground truth. This indicated that while AI-generated images can augment datasets and improve generalization, they cannot fully replace ground truth data while maintaining model performance. Integrating AI-generated images with real-world data significantly improved model performance, reduced labor-intensive data collection processes, and provided a more diverse and comprehensive dataset for training, underscoring the importance of a balanced approach to optimizing data collection protocols for wild blueberry cultivation.