AI-driven phenotyping for precision nutrient management in crops

Thesis Abstract

In precision agriculture, inefficient nutrient management undermines crop yield, increases environmental load, and elevates production costs, particularly under variable soil and climate conditions; this study addresses the gap by integrating AI-driven phenotyping with soil-plant feedback to optimize nutrient application in real time. The aim is to develop and validate a crop-phenotyping framework that links high-resolution digital phenotypes to nutrient status for site-specific management. Specific objectives are (1) to quantify leaf-level and canopy-structural phenotypes using multisensor imaging and proximal sensing to predict tissue nutrient status; (2) to develop machine learning models that map phenotypic predictors to leaf nitrogen, phosphorus, and potassium concentrations across diverse crops; (3) to formulate a decision-support algorithm for variable-rate nutrient applications that minimizes leaching while maintaining target yields; (4) to evaluate the framework under diverse edaphic and climatic scenarios; and (5) to assess economic and environmental performance relative to conventional management. A mixed-methods approach is employed in a multi-site field experiment across two growing seasons. The population comprises maize (Zea mays) and wheat (Triticum aestivum) crop stands grown on 12 experimental plots per site, spanning loamy and clay loams with different baseline fertility. The sample comprises 144 plots (12 plots × 6 treatments × 2 crops) per season, with a total of 288 plots across two seasons. Data collection instruments include high-resolution RGB and hyperspectral imaging mounted on unmanned aerial vehicles (UAVs), chlorophyll fluorescence sensors, plant height and canopy cover measurements, soil texture and nutrient assays (extractable N, P, K) from fresh soil samples, and leaf tissue analyses for nitrogen, phosphorus, and potassium using standard Kjeldahl, colorimetric, and ICP-OES methods. Ground-truthing employs destructive sampling on a stratified random subset of 24 plots per site per season for tissue nutrient validation. Methodologically, the study adopts a design-based machine learning paradigm, combining supervised learning with explainable AI (XAI) to ensure interpretability. Data are preprocessed for radiometric calibration, geometric alignment, and noise reduction; features include spectral indices (e.g., NDVI, PRI, CIGreen), textural metrics, plant architectural descriptors, and environmental covariates (soil moisture, temperature, rainfall). Regression and ensemble methods (random forest, gradient boosting, and deep neural networks) are trained to predict leaf N, P, and K concentrations. Model selection hinges on predictive accuracy (R2, RMSE) and robustness across sites, with SHAP value analyses used to interpret feature importance. A calibrated nutrient-response model then informs a variable-rate application (VRA) controller, integrated into a farm management information system (FMIS) to output site-specific fertilizer prescriptions. Temporal analyses examine phenotypic temporal dynamics to identify optimal assessment windows. Statistical analyses include multivariate ANOVA to compare management treatments, cross-validation to evaluate model generalizability, and regression diagnostics to assess assumption adherence. The theoretical lens integrates the Resource-Ability-Surplus (RAS) framework for linking plant-resource status to growth gains and the Capacity-Constraint theory of phenotyping for decision support, complemented by the Technology Acceptance Model to gauge farmer receptivity to AI-driven recommendations. Expected findings indicate that canopy and leaf-level phenotypes can predict tissue N, P, and K with R2 values exceeding 0.75 across crops, that the VRA controller reduces nitrogen losses by up to 18% and maintains yield within 3% of optimal targets, and that model interpretability via SHAP highlights nitrogen-related spectral indices as primary drivers. The study contributes to knowledge by demonstrating a scalable, AI-enabled phenotyping pipeline that translates plant- and canopy-level signals into precise nutrient management decisions, advancing integration of remote sensing, plant physiology, and agronomic practice. Implications include enhanced nutrient-use efficiency, reduced environmental impact, and improved decision support for growers under variable field conditions. Recommendations address deploying the framework across additional crops, refining cost-benefit analyses for adoption, and extending the approach to micronutrient management and soil-plant health diagnostics.