Predicción del diámetro ponderado medio de las partículas del suelo en la vereda bajo tablazo del municipio de Manizales - Colombia

Abstract

Soil structure is the arrangement and organization of the particles that constitute it, its importance lies in the fact that it allows establishing relationships with erosion and management in soil conservation systems. Its evaluation in the field from the characteristics of type and class, are complemented with the measurement of the degree of development that it presents, however, due to the quantity and cost of the samples, it is necessary to resort to predictive methods for its determination and to establish a soil transfer function. Four regression models based on supervised machine learning were evaluated: Ridge Regression, Linear Regression, Lasso Regression and Decision Tree; to predict the size of soil structural particles (Mean Weighted Diameter in mm) from properties obtained in the field together with chemical analysis of soils in the agricultural zone of the Bajo Tablazo area of the municipality of Manizales - Colombia. By training the La Floresta data set with 15 physical, chemical and biological variables; the RMSE, MSE and R2 metrics allowed evaluating the performance of the models. The trained prediction model was applied to the Bajo Tablazo data set in order to obtain a prediction of the DPM_mm. The R2 value indicated that the models are not explaining the variance of the target variable from the selected predictors. Although the application of models allows obtaining values of soil properties, in this case it was not possible to obtain this relationship. Possibly due to the low number of records and the lack of heterogeneity of the data, i.e., more diverse information from different places is required for the models to be able to generalize the behavior. It is also possible that a better preparation of the variables is necessary, here only Log+1 was applied to improve the scale of the values.