Root zone soil moisture (RZSM) plays a critical role in land-atmosphere hydrological cycles and serves as the primary water source for vegetation growth. However, the correlations between RZSM and its associated variables, including surface soil moisture (SSM), often exhibit nonlinearities that are challenging to identify and quantify using conventional statistical techniques. Therefore, this study presents a hybrid convolutional neural network (CNN)-long short-term memory neural network (LSTM)-attention (CLA) model for predicting RZSM. Owing to the scarcity of soil moisture (SM) observation data, the physical model Hydrus-1D was employed to simulate a comprehensive dataset of spatial-temporal SM. Meteorological data and moderate resolution imaging spectroradiometer vegetation characterization parameters were used as predictor variables for the training and validation of the CLA model. The results of the CLA model for SM prediction in the root zone were significantly enhanced compared with those of the traditional LSTM and CNN-LSTM models. This was particularly notable at the depth of 80-100 cm, where the fitness (R2) reached nearly 0.9298. Moreover, the root mean square error of the CLA model was reduced by 49'o and 57'o compared with those of the LSTM and CNN-LSTM models, respectively. This study demonstrates that the integration of physical modeling and deep learning methods provides a more comprehensive and accurate understanding of spatial-temporal SM variations in the root zone.