Abstract: Using observational data from a ground-based microwave radiometer at the Semi-Arid Climate and Environment Observatory (SACOL) of Lanzhou University and radiosonde data from the Yuzhong station in 2009 and 2010, a radial basis function neural network (RBFNN) algorithm was established for the inversion of air temperature, relative humidity, and water vapor density, and the application of this algorithm was explored throughout comparing the inversion results with the original products of the microwave radiometer.The results show that the maximum mean square root error is 2.72 K for air temperature, 22.32% for relative humidity, and 0.73 g·m^-3 for water vapor density, respectively, derived using the RBFNN method.The RBFNN inversion performs better than the original products of microwave radiometer at all observational heights and significantly improves profiles of air temperature at 2-10 km, relative humidity at 0-3 km, and water vapor at 1-7 km, respectively.Therefore, the RBFNN algorithm is recommended for the local inversion of meteorological variables based on a ground-based microwave radiometer.