Methodology for Selecting Stable UAV-Based Vegetation Indices for Prediction of Agronomic Variables in Maize Using a Multispectral Sensor

Plant phenotyping based on unmanned aerial vehicles still faces challenges regarding the direct correlation between spectral information with field-collected variables, due to the influence of environmental factors and the considerable variation among maize phenological stages. Therefore, the objectives of this research were: I) to evaluate the interaction of nitrogen doses and evaluation environments (phenological stages and growing seasons) and variance components for field variables and vegetation indices; II) to identify the most suitable indices according to the evaluation environments; and III) to predict field variables based on relevant vegetation indices identified through the proposed methodology. The study was conducted using a randomized complete block design with four repetitions, in which treatments consisted of six nitrogen (N) topdressing doses (0, 50, 100, 200, 300, and 400 kg ha−1) during the 2022/2023 and 2023/2024 growing seasons. Evaluations of agronomic variables and image acquisition were performed in five distinct phenological stages throughout the maize crop cycle. The data were analyzed using deviance analysis and variance components, principal component analysis (PCA), and multivariate linear modeling for the prediction of field variables. Our results demonstrated that all indices were affected by the interaction between N doses and evaluation environments (phenological stages and growing seasons). Additionally, the most reliable were EXGRaw, TGI, GNDVI, NDRE, CIRE, GVI, CVI, BNDVI, PanNDVI, SRNIRRe, SFDVI, RGBindex, NDVI, SAVI, MSAVI, and OSAVI, which showed clustering patterns according to growing season condition and phenological stage. Finally, the variables predicted using the proposed methodology achieved coefficients of determination above 0.80, except for shoot biomass and 100-grain weight. Therefore, it can be concluded that vegetation indices are influenced by the evaluated environment; however, the proposed framework based on the deduction of fixed and random effects enables the prediction of field variables with high accuracy using relatively simple models.