EVALUACIÓN DE MODELOS DE APRENDIZAJE AUTOMÁTICO PARA LA PREDICCIÓN DE BORO EN SUELOS ANDISOLES DE NARIÑO-COLOMBIA

Abstract

<p><strong>Background.</strong> The southern region of Nariño in Colombia has a strong agricultural vocation. However, it has been identified that the technology associated with the recommendation of fertilization especially in microelements in the soil is limited, which generates conflicts in production. This situation demands a different and innovative approach to address these challenges. That is why this research uses the predictive capacity of the machine learning (ML) approach. <strong>Objective.</strong> To explore the application of ML tools for the prediction of Boron levels in Andisols soils of Nariño has been explored and identifying the most efficient algorithm. <strong>Methodology.</strong> A total of 1,067 soil samples collected in various fields of five municipalities in the southern subregion of the department were used, where the supervised learning models, Random Forest (RF), K-Nearest Neighbors (K-NN), Support Vector Machine (SVM) and Naive Bayes (NB) were evaluated. <strong>Results.</strong> The results were analyzed using precision tests, kappa coefficient and confusion matrix. <strong>Conclusion.</strong> The RF algorithm demonstrated the best performance in estimating Boron levels, achieving 78% accuracy, outperforming SVM (75%), K-NN (69%) and NB (35%). <strong>Implications.</strong> These results allow for improved decision-making regarding fertilization and micronutrient management in the soil, in order to improve soil quality and therefore crop yield.</p>