PREDICTION OF LEAF MACRO AND MICRONUTRIENTS IN DIVERSE SORGHUM POPULATIONS USING PROXIMAL HYPERSPECTRAL DATA

Sorghum is a critical cereal crop for food security in semi-arid and marginal environments, where its tolerance to drought and low soil fertility supports hundreds of millions of people across Africa and South Asia. Despite the crop’s importance, the leaf macro- and micronutrient profiles of genetically diverse sorghum germplasm remain incompletely characterized, and traditional tissue analysis methods are too destructive and labor-intensive to support large scale phenotyping of breeding populations. This thesis addresses these limitations through two integrated studies using the eleven founder genotypes of the Sorghum bicolor Nested Association Mapping (SbNAM) population grown across three field environments representing two years and two different fertility levels. The first study characterized leaf nutrient variation in twelve macro- and micronutrients across the eleven SbNAM founders using mixed linear models, variance component analysis, and phenotypic plasticity assessment. Analysis of variance revealed significant genotype effects for all twelve nutrients (P<0.001), confirming broad genetic diversity in nutrient accumulation across the population. Environmental effects were significant for eleven of twelve nutrients, with calcium concentrations increasing by 48% and manganese concentrations more than doubling under low fertility conditions relative to full fertility baselines. Genotype-by-environment interactions were significant for ten nutrients, indicating genotype variation in genotype responses across environments for most traits. Phenotypic plasticity coefficients ranged from less than 5% for stable nutrients such as nitrogen to over 80% for manganese, which exhibited universal high plasticity across all genotypes. Principal component analysis confirmed that environmental effects dominated overall separation of nutrient profiles, while genotype effects were seen within-environments. These findings establish that the SbNAM founders represent a 9 phenotypically diverse population suitable for genetic studies and spectral phenotyping applications, while also identifying potential challenges for developing generalizable prediction models. The second study developed and evaluated partial least squares regression (PLSR) models for predicting the twelve foliar nutrient concentrations from proximal hyperspectral leaf reflectance collected at flowering stage using three cross validation strategies: pooled 10-fold cross-validation, within-environment 10-fold cross-validation, and leave-one-environment-out cross-validation (LOEO-CV). When calibrated on pooled multi-environment data, models achieved a moderate prediction accuracy with a mean R² of 0.398. Manganese achieved the highest accuracy (R² = 0.704), while six nutrients achieved moderate performance with R² between 0.4 and 0.7, including calcium (0.551), nitrogen (0.470), iron (0.438), boron (0.412), phosphorus (0.411), and sulfur (0.408). Within-environment cross-validation yielded intermediate performance, while LOEO-CV revealed substantial degradation of prediction models across all nutrients, with mean R² of only 0.058, demonstrating that spectral-nutrient relationships do not generalize to independent environments without recalibration. Variable importance in projection analysis showed that predictions relied overwhelmingly on the red-edge region (650–740 nm), indicating that models detected correlations with general chlorophyll status and plant vigor rather than nutrient-specific absorption features. Despite poor absolute cross-environment accuracy, genotype rankings were partially preserved for stable nutrients including nitrogen (ρ = 0.836), potassium (ρ = 0.818), sulfur (ρ = 0.809), and phosphorus (ρ = 0.745), suggesting that spectral classification of genotypes into broad accumulation categories remains viable for early-stage breeding screening even when quantitative prediction fails. The magnitude of cross-environment performance degradation was directly related to phenotypic 10 plasticity quantified in the first study, with a Spearman correlation of ρ = -0.64 (p = 0.025) between plasticity coefficients of variation and LOEO-CV R², providing a framework for identifying which nutrients are amenable to cross-environment spectral phenotyping. Together, these findings define the current capabilities and boundaries of hyperspectral nutrient phenotyping in sorghum, demonstrating potential for within-trial screening of a subset of nutrients while clarifying that cross-environment deployment requires recalibration, and establishing phenotypic stability as a key factor for trait selection in spectral phenotyping applications.