Soil moisture (SM) downscaling has been extensively investigated in recent years to improve coarse resolution of SM products. However, available methods for downscaling are generally based on pixel-to-pixel strategy, which ignores the information among pixels. Hence, a new downscaling method based on convolutional neural network (CNN) is proposed to solve the problem. Furthermore, weight layer is designed for the input and residual SM is treated as the output of CNN to improve the accuracy. This method is applied to downscale SMAP SM products (i.e., 36-km L3 SM P and 9-km L3 SM P E) from 1st January 2018 to 30th December 2018. Compared with 9-km L3 SM P E, the 9-km downscaling result is satisfactory with obtained R, RMSE, and ubRMSE values of 95.81%, 2.77%, and 2.67%, respectively. Moreover, SMAP SM products (36 and 9 km) and downscaling SM (3 and 1 km) are all validated by the in-situ data which are collected by the 109 stations of Oklahoma Mesonet (OKM) SM monitoring network.