A Comparison of Machine Learning Methods for Predicting Field Capacity and Wilting Point in Soils from Chile.

Machine learning is a multidisciplinary application combining mathematics and computer science to develop predictive models, with the intent of estimate unknown parameters. These machine-learning approaches have been successfully used in the prediction of a number of soil properties, becoming new and effective tools for the development of pedotransfer functions (PTFs).

This approach can be applied to estimate soil physical properties, including field capacity (FC) and wilting point (WP). These properties often are seen as time-consuming and expensive, as well as these measurements are commonly missing in soil surveys. For soil scientists and hydropedologists, a logical approach to handle these missing data is their estimation with PTF. This is because hydrological models based on physical properties of soils depend on these soil hydraulic properties for proper model-runs.

Throughout this study, a comprehensive comparison of some selected methods is performed upon several sets of data consisting of soils from Chile, in order to test their ability of to predict FC and WP. The performance of six models: Artificial Neural Network (ANN), Support Vector Machines (SVM), Random Forest (RF), Classification and Regression Trees (CART), Partial Least Squares (PLS) and Generalized Lineal Model (GLM) were evaluated using the Root Mean Square Error (RMSE) and the coefficient of determination (R²). These models were developed and trained based on the soil survey carried out by the Chilean government. Then, they were tested against a dataset independent of the training data. Additionally, these models were verified by comparison of predicted and actual values from external databases of Chilean soils. Preliminary results using bulk density, particle size distribution and organic matter content as predictors of FC and WP showed that ANN, RF and SVM have better performance compare to CART, GLM and PLS. This paper discusses the advantages, disadvantages and applicability of each of them as prediction tools.