Sejong University

• Addresses temporal data-resolution mismatch in greenhouses via diffusion framework • Generates biologically realistic data with environment–growth feedback model • Feedback guidance reduces errors by 69–90% across four crop growth variables • Guided feedback helps reduce the lab-to-field gap in agricultural AI applications Accurate crop growth forecasting is fundamental to precision agriculture, yet its application in real-world greenhouses faces challenges including mismatches in data acquisition frequencies. While automated environmental sensors generate high-density data streams containing millions of observations, manual growth measurements account for less than 1% of this temporal resolution. This data resolution mismatch often leads to “lab-to-field gap”, where models fail to generalise from benchmark datasets to operational settings. To address this challenge, this study introduces a feedback-guided temporal diffusion (FGTD) framework that generates synthetic growth data with data-driven causal consistency between environmental conditions and crop growth response. The FGTD framework integrates three core components: (1) a temporal diffusion model to learn from sparse growth patterns, (2) an environment-growth feedback network to enforce causal relationships, and (3) a guided reverse diffusion process to ensure biologically realistic outcomes. The framework was trained on a large-scale public dataset and validated through cross-domain testing on an independent operational tomato greenhouse dataset over the growing season with an even more severe data resolution mismatch. Ablation studies comparing FGTD against no augmentation, classical interpolation methods (linear, spline, Gaussian process regression), and standard diffusion revealed that standard diffusion improved predictions for continuous variables but failed on discrete variables like leaf count, increasing the MAE (Mean Absolute Error) by 37.3%. In contrast, the proposed feedback-guided approach achieved consistent improvements across all growth variables, reducing MAE by 74.3% for growth length, 69.2% for leaf length, and approximately 90% for both leaf width and leaf count relative to the no-augmentation baseline (leaf count bootstrap R² = 0.912 [95% CI: 0.760-0.996]; NRMSE = 0.019). Among the evaluated downstream forecasting models, XGBoost demonstrated optimal stability and accuracy when trained on the augmented dataset. These results demonstrate that environment-conditioned feedback guidance is essential for generating biologically faithful synthetic growth data, and that the FGTD framework provides a robust, scalable solution for bridging the data resolution gap toward practical deployment of AI-driven crop management systems.