UC Davis

Nitrogen (N) and nutrient management are central to sustainable grapevine production, supporting optimal yield, maximizing economic return, and reducing environmental pollution. Over the past two decades, remote sensing technologies have emerged as high potential tools for monitoring plant nutritional status. Unmanned aerial systems (UAS) and advances in computing capacity offer significant potential for nutrient management applications by enabling cost-effective acquisition of high-resolution imagery across large areas. Despite their growing popularity, traditional vegetation indices (VIs) and purely data-driven approaches face substantial limitations in generalizability, often failing to transfer across sensors, locations, and environmental conditions due to their empirical nature.This dissertation aims to develop generalizable, and interpretable methodologies for N and nutrient estimation in grapevines by leveraging physically based radiative transfer models (RTM) across multiple spatial scales. The first part of the study investigates leaf-level spectral dynamics using the PROSPECT model, evaluating the sensitivity of VIs and RTMs to biochemical and biophysical traits across different leaf ages. Comparative analysis reveals that data-driven methods achieved superior predictive performance (e.g., Gaussian Process Regression attained R² ≈ 0.78), whereas physically based models provided mechanistic interpretability grounded in biophysical parameters. A hybrid approach that combined physically based RTM-derived traits with a Random Forest Regressor yielded lower accuracy (R² ≈ 0.54) while reducing training data requirements by approximately 50%. Nevertheless, the accumulation of errors from RTM inversion and subsequent model training suggests that caution is warranted when applying the hybrid strategy.Building on these mechanistic insights, the second part of the dissertation scales the modeling framework to the canopy level through integration with the HELIOS 3D ray-tracing framework. A novel pipeline was developed in which synthetic multispectral imagery was generated under varying canopy structures, sun angles, and leaf optical properties to quantify how each factor influences overall canopy reflectance, and this imagery was used to train a deep learning model. A major milestone achieved in this study was the complete replacement of data-intensive imaging with synthetic multispectral data for model training, representing an efficient and scalable alternative to conventional data-driven approaches. The trained models demonstrated the ability to predict nutrient-related traits from real-world UAS imagery, generating spatial variability maps that identify zones requiring nutrient intervention. Although domain gaps between synthetic and real reflectance data persist the findings confirm the feasibility of using physically simulated data to train transferable models for field deployment.Collectively, this dissertation presents a hybrid modeling strategy that combines physics-based simulation and machine learning to address the fundamental trade-offs between accuracy, generalizability, and interpretability in nutrient monitoring. The resulting framework provides a strong foundation for advancing precision agriculture by enabling N estimation under physically based approach and supporting future standardization of remote sensing workflows using synthetic reference scenes.