Soil moisture is an important variable for studying global hydrological processes. Existing gridded soil moisture datasets are from data assimilation products, remote sensing, and land surface models, which are subject to considerable bias due to satellite data retrieval and modeling errors. In recent years, there has been much interest in upscaling in situ observations of ecosystem variables (e.g. evapotranspiration, gross primary productivity) to generate gridded datasets using machine learning methods. Such procedure can be similarly applied to develop upscaled gridded soil moisture datasets, which will have different error sources than existing gridded soil moisture products, and can serve as a useful alternative for data cross-checking, model evaluation, and empirical analysis. In this research, global soil moisture observations are assembled from the International Soil Moisture Network, FLUXNET, and the Canadian Global Water Futures. Random forest models are fitted between soil moisture at different depths and a variety of predictors (meteorological conditions, vegetation, soil properties, land cover, and topography), for each ecosystem type and snow/non-snow/growing/non-growing seasons. In the next step, the models will be applied with global gridded meteorological, vegetation, soil, land cover, and topography datasets to obtain global gridded long-term soil moisture product.