A Comparative Study of Various Methods for Handling Missing Data in UNSODA

Abstract

UNSODA, a free international soil database, is very popular and has been used in many fields. However, missing soil property data have limited the utility of this dataset, especially for data-driven models. Here, three machine learning-based methods, i.e., random forest (RF) regression, support vector (SVR) regression, and artificial neural network (ANN) regression, and two statistics-based methods, i.e., mean and multiple imputation (MI), were used to impute the missing soil property data, including pH, saturated hydraulic conductivity (SHC), organic matter content (OMC), porosity (PO), and particle density (PD). The missing upper depths (DU) and lower depths (DL) for the sampling locations were also imputed. Before imputing the missing values in UNSODA, a missing value simulation was performed and evaluated quantitatively. Next, nonparametric tests and multiple linear regression were performed to qualitatively evaluate the reliability of these five imputation methods. Results showed that RMSEs and MAEs of all features fluctuated within acceptable ranges. RF imputation and MI presented the lowest RMSEs and MAEs; both methods are good at explaining the variability of data. The standard error, coefficient of variance, and standard deviation decreased significantly after imputation, and there were no significant differences before and after imputation. Together, DU, pH, SHC, OMC, PO, and PD explained 91.0%, 63.9%, 88.5%, 59.4%, and 90.2% of the variation in BD using RF, SVR, ANN, mean, and MI, respectively; and this value was 99.8% when missing values were discarded. This study suggests that the RF and MI methods may be better for imputing the missing data in UNSODA.