Abstract: Atmospheric temperature and humidity profiles are critical input parameters for satellite remote sensing retrieval, numerical weather prediction, and climate model simulations. The GNOS occultation detector onboard the Fengyun series satellites has achieved continuous global observation of atmospheric parameters, providing important data support for weather forecasting and extreme weather monitoring. However, the detection accuracy in the near-surface boundary layer is affected by underlying surface conditions and needs improvement. Radiosonde observations have high accuracy but limited temporal resolution with only twice-daily measurements, making it difficult to support continuous satellite validation. Ground-based microwave radiometers achieve continuous retrieval of temperature and humidity profiles by receiving atmospheric downwelling radiation brightness temperature, without underlying surface limitations, providing an important approach for satellite data comparison and validation. However, traditional retrieval methods have limited accuracy in the boundary layer. This paper presents an improved Transformer design for atmospheric retrieval tasks. Using radiosonde measurements from the Beijing South Suburban Observatory from 2020 to 2025 and MonoRTM radiative transfer model forward simulations to construct training datasets, a ground-based retrieval model based on the self-attention mechanism was established and compared with linear regression and BP neural networks. Results show that the Transformer retrieval accuracy is significantly improved compared to linear regression and BP neural networks, with temperature RMSE reduced by 29.7% and 10.1% respectively, and absolute humidity RMSE reduced by 26.8% and 4.6% respectively. Further comparison and validation of the FY-3E GNOS-II occultation product in the 0-10 km troposphere using radiosonde as truth show that the Transformer retrieval temperature and absolute humidity RMSE are improved by 61.1% and 71.3% respectively compared to the occultation product. This method combines the high temporal resolution of microwave radiometers with retrieval accuracy approaching that of radiosondes, serving as a bridge between sparse radiosonde and continuous satellite observations, providing dynamic validation benchmarks for Fengyun satellite boundary layer products, and supporting the development of satellite retrieval algorithm optimization and ground-satellite collaborative observation systems.