Abstract

Chile, which accounts for 27% of global cherry exports (USD 2.26 billion annually), faces a critical industry challenge in crop health monitoring. While automated sensors monitor environmental variables, phytosanitary diagnosis still relies on manual visual inspection, leading to detection errors and delays. Given this reality and the growing use of AI models in agriculture, our study quantifies the theory–practice gap through comparative evaluation of three transfer learning architectures (namely, VGG16, ResNet50, and EfficientNetB0) for automated disease identification in cherry leaves under both controlled and real-world orchard conditions. Our analysis reveals that excellent laboratory performance does not guarantee operational effectiveness: while two of the three models exceeded 97% controlled validation accuracy, their field performance degraded significantly, reaching only 52% in the best-case scenario (ResNet50). These findings identify a major risk in agricultural transfer learning applications: strong laboratory performance does not ensure real-world effectiveness, creating unwarranted confidence in model performance under real conditions that may compromise crop health management. © 2025 by the authors.