Abstract
Maintaining crop health is essential for global food security, yet traditional plant monitoring methods based on manual inspection are labor-intensive and often inadequate for early detection of stressors and diseases, and insufficient for timely, proactive interventions. To address this challenge, we propose a deep learning-based framework for expert-level, spatiotemporal plant health assessment using sequential RGB images. Our method categorizes plant health into five levels, ranging from very poor to optimal, based on visual and morphological indicators observed throughout the cultivation cycle. To validate the approach, we collected a custom dataset of 12,119 annotated images from 200 tomato plants across three varieties, grown in semi-open greenhouses over multiple cultivation seasons within one year. The framework leverages state-of-the-art CNN and transformer architectures to produce accurate, stage-specific health predictions. These predictions closely align with expert annotations, demonstrating the model's reliability in tracking plant health progression. In addition, the system enables the generation of dynamic cultivation maps for continuous monitoring and early intervention, supporting data-driven crop management. Overall, the results highlight the potential of this framework to advance precision agriculture through scalable, automated plant health monitoring, guided by an understanding of key visual indicators and stressors affecting crop health throughout the cultivation period.