Guiafertil: software de análise exploratória e predição do limiar de produtividade da cultura de soja com técnicas de inteligência artificial

Abstract

Soybean cultivation plays a fundamental role in Brazilian agribusiness, requiring increasingly efficient management practices to ensure high levels of productivity and sustainability. However, the criteria currently used by agronomists to recommend fertilizer and soil amendment rates exhibit considerable variability, mainly due to the lack of regional calibration studies that simultaneously consider specific soil attributes, climatic conditions, altitude, planting systems, and other variables that influence crop development. This limitation hinders technical decision-making in the field and reduces the accuracy of agronomic recommendations. In this context, understanding the relationship between these factors and crop productivity becomes essential to enable more robust estimates and support management decisions. Advances in Artificial Intelligence (AI) techniques offer new opportunities to model such relationships and identify complex patterns that are often difficult to detect using traditional approaches. Therefore, this study presents the development of Guiafertil, a computational tool designed for exploratory analysis of agronomic data and soybean yield threshold prediction. Guiafertil integrates a database management system, data preprocessing routines, statistical analysis, graphical visualization, and machine learning algorithms, including Linear Regression, Support Vector Regression (SVR), Random Forest, and Artificial Neural Networks. Using real-world data obtained from the “Champion Cases” database of the Brazilian Soybean Strategic Committee (CESB), the system enables the evaluation of soil, climate, and management attributes, model training, performance analysis, and productivity estimation based on practical scenarios, such as different fertilization rates. Among the main results, the Multi-Layer Perceptron (MLP) Neural Network achieved the best predictive performance, reaching an R² value of 0.56 for soybean yield prediction bags/ha, whereas Linear Regression showed significant predictive limitations, obtaining an R² value of -0.60. Furthermore, the case study demonstrated that adjustments in fertilizer application rates affect production costs but have limited impact on predicted productivity, highlighting the importance of more accurate decision-making processes. The software quality and usability were evaluated by agronomists, who assigned average scores above 9.0 for most of the assessed criteria. Therefore, Guiafertil proved to be a promising tool for supporting agronomic recommendations, contributing to more efficient use of agricultural inputs, cost reduction, and greater sustainability in soybean production. Additionally, this work opens new opportunities for future research involving other crops, expansion of the database, and integration with precision agriculture technologies.