Deriving surface skin temperature and atmospheric profiles of temperature and water vapor from HIRS measurement using a neural network technique

A neural network technique is applied to develop an algorithm to derive surface skin temperature, atmospheric temperature, and water vapor from HIRS measurement. The neural network training dataset is constructed by combining a set of diverse temperature and water vapor profiles with corresponding HIRS channel brightness temperatures simulated by a radiative transfer model. By applying the retrieval algorithm to an independent validation dataset, the result shows that the root-mean-square (rms) error of the surface skin temperature retrieval is 0.7 C. For the atmospheric temperature profile retrieval, the rms errors are 1.5-2.4 C at the near-surface levels, about 1.3 C in the mid-troposphere, and 1.5-2.1 C around the tropopause. For the water vapor profile retrieval, the rms error at 1000 mb is 2.1 g/kg. The rms errors decrease steadily to 1.8 g/kg at 800 mb, 1.2 g/kg at 700 mb, 0.7 at 600 mb, and less than 0.4 g/kg above 500 mb. The retrieval algorithm will be applied to a time series of global HIRS data to study the temperature and water vapor variability.