Abstract

The soil suitable for cultivation is influenced by complex physical, biological, and chemical interactions that directly affect its quality, as well as conventional agricultural practices such as monoculture, tillage, and fertilization, making it a determining factor for the success of banana productivity. Therefore, it is important to evaluate which variables significantly influence soil productivity. This study employed a dataset corresponding to 17 explanatory variables (pH, electrical conductivity, cation exchange capacity, organic matter, macro and microelements) retrieved from 6 banana fields (A1 – A6) with unidentified production levels, from which 60 sampling points were selected for soil characterization. The main objective was to present and evaluate logistic regression models of machine learning that encompass soil chemical properties and identify those that best explain banana productivity variability. Additionally, the study included a productivity index as a response variable for the models, based on values on the morphological dimensions of banana plants in their respective area. Several regression models were estimated using stepwise regression method, and subsequently, the prediction quality of the machine learning models obtained was analyzed. The final model included 5 variables explaining the Productivity Index with an R2 greater than 0.80, therefore, these variables are proposed for implementation in solving agricultural issues. Through this approach, we aim to provide important sources of information for farmers, researchers, and decision-makers, contributing to the optimization of agricultural practices and the promotion of more efficient and sustainable cultivation systems.