A Neural Network Approach to Improve the Vertical Resolution of Atmospheric Temperature Profiles From Geostationary Satellites

Abstract

Tropospheric temperature measurements at high temporal, spatial, and vertical resolutions are required for many meteorological studies. Radiosonde and Global Positioning System radio occultation (GPSRO) observations have very high vertical resolutions but poor in spatial and temporal coverage. Although the sounders on geostationary satellites can provide high temporal and spatial resolutions, their vertical resolution is poor. In this letter, we proposed a method to increase the vertical resolution of tropospheric temperature profiles obtained from geostationary satellite observations based on an artificial neural network (ANN) approach so that high-resolution temperature profiles are available in all four dimensions. We simulated the pressure levels of the forthcoming Indian National Satellite System (INSAT) 3-D temperature measurements from 950 to 100 hPa using 1-D variational temperature profiles of the Constellation Observing System for Meteorology Ionosphere and Climate (COSMIC). We used these low-resolution simulated profiles as the predictors and the high-resolution GPSRO COSMIC profiles as predictants. The data during 2007 and 2008 were used to develop the model, and the data during 2009 were used for validation. The correlation coefficient of greater than 0.94 is observed throughout the pressure levels for all the three data sets. The root-mean-square differences of training, selection, and validation sets are 0.43, 0.46, and 0.51, respectively. A scatter index of less than 0.002 for all the three data sets indicates the accuracy of the estimations.