A neural network technique is applied to retrieve temperature and water vapor profiles from satellite sounder HIRS measurement. The training dataset is constructed with a diverse sample of atmospheric profiles from global model reanalysis data and simulated HIRS brightness temperatures by a radiative transfer model. Previous work has generated inter-satellite calibrated, cloud-cleared, and limb-corrected HIRS channel brightness temperature data. By applying the neural network retrieval scheme to the HIRS dataset, a twenty-year time series (1980-1999) of temperature and water vapor profiles is constructed at 2.5 degree latitude/longitude grids. With the satellite sounder measurement, the retrieval scheme is able to provide temperature and water vapor profile data at a global scale, covering areas that may not be routinely observed by radiosondes. The long-term variation of temperature and water vapor and the global distribution of these variables at different levels of the atmosphere will be shown.