Estimating the soil water retention curve: Comparison of multiple nonlinear regression approach and random forest data mining technique

This study evaluates the performance of the random forest (RF) method on the prediction of the soil water retention curve (SWRC) and compares its performance with those of nonlinear regression (NLR) and Rosetta-based pedotransfer functions (PTFs), which has not been reported so far. Fifteen RF and NLR-based PTFs were constructed using readily-available soil properties for 223 soil samples from Iran. The general performance of RF and NLR-based PTFs was quantified by the integral root mean square error (IRMSE), Akaike's information criterion (AIC) and coefficient of determination (R-2). The results showed that the accuracy of the RF-based PTFs was significantly (P < 0.05) better than the NLR-based PTFs, and that the reliability of the NLR-based PTFs was significantly (P < 0.01) better than the RF-based PTFs and all of the Rosetta-based PTFs. The average values of the IRMSE, AIC and R-2 of the RF method were 0.041 cm(3) cm(-3), -16997.7, and 0.987, and 0.053 cm(3) cm(-3), -15547.5, and 0.981 for the training and testing steps of all PTFs, respectively, whereas the values for the NLR method were 0.046 cm(3) cm(-3), - 16616.4, and 0.984, and 0.048 cm(3) cm(-3), -16355.6, and 0.983 for the training and testing steps, respectively. The PTF5 of the RF and NLR methods, with inputs of sand and clay contents, bulk density, and the water content at field capacity and permanent wilting point, had the greatest R-2 values (0.987 and 0.989, respectively), and the lowest IRMSE values (0.039 and 0.032 cm(3) cm(-3), respectively) compared to other PTFs for the testing step. Overall, the RF method had less reliability for the prediction of the SWRC compared to the NLR method due to overprediction, uncertainty of determination of forest scale and instability in the testing step. These findings could provide the scientific basis for further research on the RF method.