Evaluation of machine learning algorithms in the prediction of hydraulic conductivity and soil moisture at the Brazilian Savannah

Abstract

The Brazilian Savannah (Cerrado biome) is the the main agricultural region in Brazil. The Cerrado has experienced a growing intensification of agriculture and an increase in disputes over water use, highlighting the need to establish strategies to reduce the water withdrawn from waterbodies, mainly through irrigation. The lack of data at a proper scale on soil hydraulic properties in the region brings uncertainties to the process of water resources management. Obtaining these data, however, is difficult and costly, thus, opening the opportunity for the use of Pedotransfer Functions (PTFs). Various methods can be used to obtain PTFs, but currently, machine learning techniques are gaining strength. In this context, it becomes important to evaluate the quality of machine learning algorithms in predicting PTFs. The present work aimed to evaluate the performance of machine learning algorithms in the prediction of saturated soil hydraulic conductivity (Ks) and soil moisture at tensions of 0, 6, 10, 33, 100, and 1500 kPa for the Cerrado Biome. Four machine learning algorithms were tested: Multiple Adaptive Regression Splines (MARS), Random Forest (RF), Support Vector Regression (SVR), and K Nearest Neighbors (KNN). Four combinations of soil data were evaluated and the predictor variables used in each set were different. In set A1, the following variables were used: sand (Sa), silt (Si), and clay (Cl) contents; in set A2: Sa, Si, Cl, and bulk density (BD); in set A3: Sa, Si, Cl, BD, particle density (Dp), total porosity (Pt), microporosity (Micro), and macroporosity (Macro); and in set A4: Sa, Si, Cl, BD, Dp, Pt, Micro, Macro, soil moisture at field capacity (θ10), and soil moisture at permanent wilting point (θ1500). The set A4 along with the RF and SVR models had the best performances in the prediction of Ks. As for soil moisture, the RF, SVR, and MARS models showed the best performances with low RMSE and ME values, and R2 above 0.8 using predictor sets A3 and A4.