Fenotipagem por Vant e aprendizado de máquina para predição da produtividade de milho em ambientes contrastantes de nitrogênio em região semiárida

Abstract

High-throughput phenotyping has contributed substantially to advances in agriculture and plant breeding programs. By using unmanned aerial vehicles (UAVs), its application has enabled the rapid acquisition of phenotypic data with high reliability and spatial resolution, promoting its adoption across diverse agricultural contexts. Despite these advances, several methodological limitations remain insufficiently explored, particularly when this approach is applied under conditions of environmental stress and high spectral homogeneity. This study aims to evaluate the accuracy of four predictive models, using regression and machine learning approaches to estimate the grain yield of maize half-sib progenies grown under contrasting nitrogen conditions, based on multispectral images obtained by UAV. The experiments were conducted during the 2024 growing season in two experimental areas located in the municipalities of Nossa Senhora da Glória and Graccho Cardoso, Sergipe, Brazil. A randomized block design with split plots and three replications was adopted, in which nitrogen levels (high and low) were assigned to the main plots and the half-sib progenies to the subplots. Multispectral drone images were acquired during three flights at heights of 60 and 80m throughout the crop cycle, corresponding to different phenological stages. Fourteen multispectral vegetation indices associated with canopy structure, chlorophyll content, and plant nutritional status were calculated. Four predictive approaches were evaluated: a classical linear model, a penalized regression model (LASSO), and two decision tree-based algorithms, Conditional Forest (Cforest) and Gradient Boosting Machine (GBM). Model performance was evaluated using the root mean square error (RMSE) and the mean absolute error (MAE), in addition to diagnostic analyses of the linear model assumptions and assessments of variable importance and SHAP values for the GBM and Cforest models. In both experimental areas and nitrogen conditions, the Cforest and GBM models consistently showed higher predictive accuracy than the linear approaches. Overall, images acquired at 60m resulted in superior model performance, highlighting the importance of higher spatial resolution for increased sensitivity to physiological variability within the canopy. Analyses of variable importance and SHAP values indicated that predictive contributions were concentrated in a limited subset of vegetation indices, particularly those sensitive to canopy structure and chlorophyll content. Furthermore, distinct predictive responses were observed among genotypes, depending on the modeling approach and environmental conditions. These results suggest that the acquisition and processing of spectral data can be optimized by prioritizing specific vegetation indices and strategically selected phenological stages. For predicting maize grain yield, the combined use of the Cforest and GBM models with vegetation indices such as TCARI, MCARI, and MSAVI2 is recommended. To increase the robustness and generalizability of these findings, future studies should evaluate the performance of such indices across a broader range of nitrogen levels, genotypes, and edaphoclimatic conditions.