Remote sensing holds significant advantages in large-scale soil moisture (SM) monitoring, providing numerous satellite SM products with valuable spatio-temporal insights and timely data updates. However, some popularly used satellite SM products, such as the European Space Agency Climate Change Initiative (ESA CCI) SM product, suffer from substantial data gaps. These gaps severely hamper its utility in large-scale meteorological and hydrological applications. To address these limitations, this study introduces an innovative gap-filling approach for reconstructing daily SM time series using the ESA CCI SM product. Our method employs a hierarchical framework that integrates the k-means clustering algorithm with a self-attention filling model and is applied to China. Through systematic division into four sub-regions based on climatic differences, specialized deep learning models are individually trained to fill gaps. The proposed method was validated using simulated gaps (real data as reference) and extended triple collocation analysis (assumed ground data as reference), along with comparison to four existing SM datasets. Results show that the reconstructed data have high correlation (R > 0.90) and low error (RMSE < 0.026 m3/m(-3)(-|-)) across the four regions in simulated gaps. Further analysis suggests that the reconstructed data's accuracy is comparable to or exceeds that of the original ESA CCI data, with a notable improvement of approximately 3 % in R accuracy during the summer season. These results emphasize the effectiveness of the proposed framework, making a promising contribution to the advancement of SM monitoring and environmental research.