Aplicação de técnicas de inteligência artificial e dados de sensoriamento remoto por imagem digital para predição da produtividade de soja (Glycine Max L.) em região do Cerrado

Abstract

Soybean (Glycine max L.) is one of the most important crops in the world's agricultural production and essential in the context of food security. The global demand for food is growing more and more, and the adoption of new technologies that allow this production to increase in a sustainable way is necessary. In this context, the objective of this work was to verify the potential of using artificial intelligence models to estimate soybean yields based on remote sensing (RS) data obtained through aerial and orbital platforms. For this study, an area of approximately 100 hectares cultivated with soybeans during the 2021/2022 harvest was used, which was subdivided into experimental plots, totaling a sample group of 5011 plots intended for the study. The spectral response of the crop canopy at different phenological stages was measured, and finally a yield survey was conducted with a grain harvester. These spectral data were combined, and together with vegetation indexes, used to predict soybean yields through Artificial Neural Network (ANN) models. Two ANN architectures were modeled, the first using data from the Planet Scope Satellite captured at five different phenological stages of the soybean crop, and the second from UAV data at the R5 stage. The fits of the models were evaluated for the plot of test data using the statistical indices of Pearson's correlation coefficient (r), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and coefficient of determination (R²). For the prediction model based on satellite images, the best results found were in the R5 vegetative stage, with correlation values of 0.71, mean absolute error of 175.45 kg ha-1 representing a percentage error of 3.68% and R² of 0.5. As for modeling from UAV data, Pearson correlation values of 0.72, 172.94 kg ha-1 for MAE, representing an MAPE of 3.66%, and R² of 0.51, were found. Both models showed no tendency to cluster errors spatially, having a global Moran's index of around 0. The results indicate that the use of both satellite and UAV imagery, modeled by ANNs, can provide accurate and robust prediction of soybean yields.