Atmospheric humidity is a fundamental parameter in weather forecasting and atmospheric science, playing a critical role in weather analysis and numerical simulation. Providing specific humidity profiles with broad spatial coverage and high accuracy remains a key challenge for improving the performance of numerical weather prediction models. In this study, we propose a deep learning approach for retrieving atmospheric specific humidity profiles at multiple pressure levels across China, using microwave humidity sounder (MHS) data and a UNet3+neural network against ERA-5 reanalysis data from the European Center for Medium-Range Weather Forecasts (ECMWF). This method effectively mitigates the ill-posedness and uncertainty commonly encountered in traditional quantitative remote sensing retrievals, enabling robust and accurate humidity profile estimations across long-term series and multiple atmospheric layers. Experiments were conducted at the 700 hPa level, with data from 2011—2019 used for training, 2020 for validation, and 2021 for testing. Results show a slight underestimation of specific humidity compared with ERA-5, though seasonal differences in retrieval error were not statistically significant. Spatially, larger retrieval errors were observed in southern China and over land surfaces, while lower errors occurred in northern regions and over oceans. The root mean square error (RMSE) remained below 1.3 g-kg in all seasons, with the lowest average RMSE of 1.15 g-kg in winter. Temporal correlation coefficients exceeded 0.9, with an annual mean of 0.92, indicating strong spatial and temporal consistency with ERA-5. Further analysis was conducted across pressure levels from 300 hPa to 1 000 hPa. The retrieved specific humidity showed good agreement with ERA-5 in spatial patterns, with RMSE values across all levels remaining below 1.53 g-kg and correlation coefficients above 0.9. The retrieval accuracy improved with decreasing altitude, showing better agreement near the surface. Comparisons with radiosonde data confirmed these results, with an average annual RMSE of 0.91 g-kg from 300 hPa to 1 000 hPa. The inversion results were slightly lower than radiosonde observations overall, particularly above 700, while at near-surface levels (e. g., 1 000 hPa) a notable RMSE reduction of approximately 0.6 g-kg was observed. These findings demonstrate the effectiveness and high accuracy of the proposed deep learning-based inversion method for retrieving atmospheric specific humidity profiles from satellite microwave data.