Soil moisture (SM) data assimilation is often challenged by temporally sparse and irregular observations, particularly in data-scarce regions such as the Tibetan Plateau, where observational gaps and uneven sampling frequencies can substantially degrade assimilation performance. Conventional approaches, including the ensemble Kalman filter (EnKF), require repeated covariance estimation and matrix inversion, leading to considerable computational cost and reduced effectiveness under intermittent observation conditions. To address these limitations, we develop a deep learning-augmented SM data assimilation framework that explicitly accounts for temporal irregularity while improving computational efficiency. The framework consists of two complementary components. First, a fully connected neural network (FCNN) is trained to emulate the nonlinear analysis-update process of EnKF, thereby reducing computational overhead by avoiding explicit covariance calculations. Second, a gated recurrent unit with decay mechanism (GRU-D) is introduced to capture temporal dependencies from irregularly sampled observations through explicit incorporation of elapsed-time information. The framework is evaluated using observations from eight sites across the Tibetan Plateau. Results show that the proposed approach consistently outperforms the conventional EnKF under sparse and irregular observation scenarios. Across all sites, the framework reduces RMSE by more than 60 % and increases correlation coefficients by over 80 % relative to EnKF, while reducing computational time by more than 30 %. The largest improvements are observed under highly irregular observation conditions, demonstrating the importance of explicitly accounting for temporal sparsity in data assimilation. These findings highlight the potential of deep learning to enhance both the robustness and efficiency of soil moisture data assimilation and provide a practical solution for hydrological state estimation in observation-limited environments.