To accelerate plant breeding genetic gain, spatial heterogeneity must be considered. Previously, design randomizations and spatial corrections have increased understanding of genotypic, spatial, and residual effects in field experiments. This study proposes a two-stage approach for improving agronomic trait genomic prediction (GP) using high-throughput phenotyping (HTP) via unoccupied aerial vehicle (UAV) imagery. The normalized difference vegetation index (NDVI) is measured using a multi-spectral MicaSense camera and ImageBreed. The first stage separates additive genetic effects from local environmental effects (LEE) present in the NDVI throughout the growing season. Considered NDVI LEE (NLEE) are spatial effects from univariate/multivariate two-dimensional splines (2DSpl) and separable autoregressive (AR1) models, as well as permanent environment (PE) effects from random regression models (RR). The second stage leverages the NLEE within genomic best linear unbiased prediction (GBLUP) in two distinct implementations, either modelling an empirical plot-to-plot covariance (L) for random effects or modelling fixed effects (FE). Testing on Genomes-to-Fields (G2F) hybrid maize (Zea mays) field experiments in 2017, 2019, and 2020 for grain yield (GY), grain moisture (GM), and ear height (EH) improves heritability and model fit equally-or-greater than spatial corrections; however, genotypic effect estimation across replicates is not significantly improved. Electrical conductance (EC), elevation, and curvature from a 2019 soil survey significantly improve GP model fit, but less than NLEE. Soil EC and curvature are most correlated to univariate 2DSpl NLEE. Defining L significantly improves genomic heritability and model fit more than setting FE, and RR NLEE can most significantly improve GP for GY and GM.