RF_global_SSM_2000-2019_0.25degree_V1

Most of long-term gridded surface soil moisture (SSM) products are from multi-satellite data fusion/blending, with the land surface model simulated SSM as the climatology reference. The inherent biases of different models, when compared with in-situ soil moisture observations, imply that the use of different land surface models will lead to different spatio-temporal patterns of soil moisture. International Soil Moisture Network (ISMN) delivers quality controlled long-term in-situ soil moisture observations, which provides the opportunity for producing global SSM with machine learning approach. In this study, the Random Forest (RF) model was trained to compute SSM from relevant feature variables (i.e., land surface temperature, vegetation indices, soil texture and geographical information) and precipitation based on the in-situ soil moisture data of the ISMN. The results indicated that the trained RF model is able to capture the highly nonlinear relationship between the SSM and the land surface variables, and can be used to generate a long-term in-situ observation constrained SSM at the global scale (0.25-degrees). The results of RF model show an RMSE of 0.05 m³/m³ and a correlation coefficient of 0.9. The calculated impurity-based feature importance indicates that the Antecedent Precipitation Index affects the predicted soil moisture most. The geographical coordinates also influence the estimation significantly (i.e., RMSE was reduced 0.03 m³/m³ after considering geographical coordinates), followed by land surface temperature, vegetation indices and soil texture. The spatio-temporal pattern of RF predicted SSM was compared with the European Space Agency Climate Change Initiative soil moisture product, using both time-longitude and -latitude diagrams. The results indicate that the RF-predicted SSM captures not only the spatial distribution, but also the daily, seasonal, and annual variabilities of SSM globally. <br>