Soil moisture is an important variable linking the atmosphere and the terrestrial ecosystems. However, long-term satellite monitoring of surface soil moisture is still lacking at global scale. In this study, we conducted data fusion of up to 11 microwave remote sensing-based soil moisture products through a set of neural networks, with SMAP soil moisture data applied as the fundamental training target. The training efficiency proves to be high (R2 =0.95) due to the selection of 9 quality impact factors of microwave soil moisture products and the elaborate organization structure of multiple different neural networks (5 rounds of simulation; 8 substeps; 74 independent neural networks; and >106 regional subnetworks). We achieved global satellite monitoring of surface soil moisture during 2003~2018 at 0.1° resolution. This new dataset, once validated against the International Soil Moisture Network (ISMN) records, is found superior to the existing products (ASCAT-SWI, GLDAS Noah, ERA5-Land, CCI/ECV and GLEAM), and is applicable to studying both the spatial and temporal patterns. It suggests an increase in global surface soil moisture, and reveals that the surface moisture decline on rainless days is highest in summers over the low-latitudes but highest in winters in most mid-latitude areas. Notably, the error propagation with the extension of the simulation period to the past is well controlled, indicating that the fusion algorithm will be more meaningful in future when more advanced sensors are in operation.