To obtain high-precision soil moisture data in basins, three types of soil moisture data from SMAP, AMSR2, and CLDAS were integrated. Considering the effects of precipitation, normalized difference vegetation index, slope, and other factors on soil moisture, a convolutional neural network (CNN) downscaling model for soil moisture that can consider the spatial neighborhood relationship of auxiliary factors was established based on CNN, and the soil moisture data with a spatial resolution of 1 km were obtained. The model was applied to the basin from Pushi County to Wuqiangxi Dam site in Hunan Province, and the results show that the spatial distribution characteristics of soil moisture data before and after downscaling are consistent, and the soil moisture data after downscaling can correctly reflect the response of soil moisture to flood events and add more spatial distribution details. Compared with the measured soil moisture data at soil moisture stations, the mean bias, mean absolute error, and root mean square error of the downscaled soil moisture are -0.061, 0.086, and 0.099 cm3/cm3. The CNN downscaling model also shows better stability in comparison with the results of the random forest downscaling model.