Temporal and spatial variations of soil moisture are crucial to agriculture, climate, water resources, and other aspects, but the current data resolution is not enough to meet the needs of small- and medium-scale monitoring. In this study, the super-resolution model is applied to soil water downscaling for the first time. To meet the requirements of lightweight and rapid inference in soil water downscaling tasks, two lightweight soil water super-resolution algorithms, ACSRGAN-1 and ACSRGAN-2, are proposed. These algorithms make use of structural reparameterization methods found in ACNet to improve SRGAN. Among them, ACSRAGN-1 uses ACBlock stacking to replace the backbone network in the original SRGAN generation network and uses structural reparameterization to effectively reduce the number of parameters in the network model. ACSRAGN-2 uses ACBlock instead of SRGAN to generate convolution operations in the network. The experimental results on the soil moisture dataset of the Black River Basin show that the accuracy of the two improved methods reaches 34.92dB and 35.22dB, and the parameter number of ACSRGAN-1 decreases by 21.59% compared with SRGAN.