PREDIÇÃO DA PRODUTIVIDADE DE LINHAGENS ENDOGÂMICAS DE MILHO UTILIZANDO ÍNDICES DE VEGETAÇÃO E APRENDIZAGEM DE MÁQUINA

Abstract

Predicting the productivity of inbred corn is crucial in modern agriculture because it helps genetic improvement companies make decisions, supports crop management during research, and provides more agile monitoring of trial plots. However, this task is challenging, since the association of spectral data with an ideal type of Machine Learning (ML) algorithm to predict crop productivity is not yet well established. Thus, the objective of this work was to predict corn yield using spectral bands and grain yield of two harvests as base data to test the performance of six different ML algorithms. The experiment was implemented in the 2023 and 2023/2024 harvests at the Federal University of Mato Grosso do Sul (UFMS), Chapadão do Sul Campus. The experimental design adopted was that of increased blocks, with 100 inbred corn lines and 10 open-pollinated varieties used as controls. The experiment was implemented in the 2023 and 2023/2024 harvests at the Federal University of Mato Grosso do Sul (UFMS), Chapadão do Sul Campus. The experimental design adopted was that of increased blocks, with 100 inbred corn lines and 10 open-pollinated varieties used as controls. Spectral image collection was performed with the aid of a remotely piloted aircraft (RPA) and subjected to static analysis. Reflectance values were obtained in the following spectral bands (SBs): blue (475 nm), red (660 nm), green (550 nm), Rededge (735 nm) and NIR (790 nm). The images were processed and 42 vegetation indices (VIs) were calculated. The data were subjected to ML analyses, testing three types of inputs: using only IVs, using only SBs and using IVs+SBs as input configuration. The machine learning models evaluated were: Artificial neural networks (ANN), REPTree decision tree (DT) and M5P; Random Forest (RF); Support vector machine (SVM) and ZeroR (used as a control model). The efficiency of the prediction models was evaluated through the metrics correlation coefficients (r), mean absolute error (MAE) and root mean square error (RMSE). The algorithms RNA, DT, M5P, RF and SVM had accuracy higher than 0.6 but M5P and SVM obtained the highest accuracies. There was no significant difference in accuracy between the inputs tested, allowing the use of only SB.