Predição de produtividade de sementes de soja usando atributos do solo e aprendizagem de máquina

Abstract

There are several studies that sought to understand how the chemical attributes of the soil influence the yield of soybeans. These researches used, for the most part, geostatistical techniques and multivariate analysis to demonstrate the effect of variables such as organic matter content, CEC, clay on the productive components and physiological performance of soybean. The objective of the research was to test different chemical attributes as input in machine learning models to estimate soybean seed productivity. The present work was carried out using the database (BD) of soil and grain yield of the ATTO Sementes sowing for the soybean crop grown in the 2020/2021 season. The evaluated soil attributes were pH, cation exchange capacity (CEC), base saturation (SB), clay content, organic matter (OM). Seed yield was obtained through the generation of a harvest map, by the integrated JDLink system of the John Deere S790 Harvester. Pearson's correlation was performed to verify the interrelationship between the analyzed soil variables and productivity. The data were subjected to machine learning analysis (artificial neural networks, linear regression, M5P, REPTree, random forest and support vector machine). Six different configurations for the algorithms were tested: pH, CTC, V%, altitude, clay and all information together. As an output variable (output) of the algorithms, soy bean productivity was used. The use of all soil variables (pH, CEC, SB, clay content and MO) associated with the random forest machine learning model makes it possible to predict soybean seed yield with high precision.