GeoAI framework for crop nutrient and above ground biomass estimation using UAV multispectral remote sensing and machine learning across diverse Virginia environments
Files
TR Number
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Accurate, timely, and spatially explicit estimation of crop biomass and nutrient status is essential for precision nutrient management in winter wheat (Triticum aestivum L.). Conventional assessment relies on destructive tissue sampling and laboratory analysis, which is labor-intensive, costly, and unable to capture field-scale spatial variability within the time frame required for in-season decisions. This dissertation addresses these limitations by developing a consistent unmanned aerial vehicle (UAV) multispectral remote sensing and machine learning (ML) framework for estimating aboveground biomass (AGB), grain yield, and the uptake of nitrogen (N) and phosphorus (P). The framework integrates spectral vegetation indices (VIs), gray-level co-occurrence matrix (GLCM) texture, normalized difference texture indices (NDTI), and ancillary variables within multiple linear regression (MLR) and random forest (RF) models. It was evaluated across 176 plots at three contrasting Virginia field sites during the 2024-2025 growing season. The research began with a systematic map of remote sensing algorithms for N and P estimation. This synthesis revealed a pronounced technological divergence: N estimation has matured into a reliable tool, whereas P estimation remains a substantial and largely unresolved challenge, attempted in only a small minority of studies. These findings defined the gaps that the empirical chapters were designed to address. The first empirical study estimated AGB and grain yield. Random forest combined with NDTI cross-band texture achieved its highest accuracy at the tillering stage (R2= 0.84; RPD=2.57), establishing tillering as the optimal imaging window, with accuracy declining through later stages as the canopy saturated and senesced. Direct yield prediction proved unreliable due to lodging-induced harvest losses but applying site-specific harvest index (HI) values to AGB maps produced credible spatially continuous yield estimates. The second empirical study estimated P uptake. Estimation was most reliable at tillering, where texture and cross-band features captured P-driven structural variation that spectral indices alone could not represent (RF R2=0.86; RPD=2.77), while remaining challenging during canopy transition at booting. The resulting P uptake maps resolved within-field variability and differences attributable to contrasting fertilization histories. The third empirical study estimated plant N accumulation. Reliable predictions were obtained at tillering (RF R2=0.84; RPD=2.67) and at a useful level at booting, whereas at heading poor prediction was likely due to N values convergence at canopy closure. Notably, ground-measured ancillary variables, including plant height, chlorophyll readings, and soil chemistry, added little once imagery features were available, supporting a workflow based on UAV imagery alone. These studies demonstrate that a single, consistent UAV multispectral framework can estimate AGB, nutrient status, and yield in winter wheat. The estimation accuracy is governed in a systematic and interpretable manner by growth stage, feature type, and the nutrient concerned. Tillering was consistently the most informative stage, texture features and RF contributed most strongly to the early season, and N proved far more tractable than P. By clarifying when and why UAV-based estimation succeeds, and by establishing a largely imagery-based workflow validated across diverse environments, this research provides both new scientific understanding and practical tools for precision nutrient management in Mid-Atlantic winter wheat production.