Abstract
A new method is presented of statistical retrieval of humidity profiles based on measurements of surface temperature ξ1, surface dewpoint ξ2, and integrated water vapor ξ3. In this method the retrieved values of humidity depend nonlinearly on predictors ξ1,2,3. A self-training algorithm was developed to obtain the values of parameters that enter into the retrieval algorithm. The data from two years of measurements in eight different locations were used for training. The method was applied to an independent dataset (including nonmonotonic profiles) of one month of surface measurements and integrated water vapor obtained from microwave radiometers. Three constraints were imposed: 1) the integrated retrieved humidity profiles had to be equal to the measured values ξ3, 2) the retrieved surface humidity had to coincide with the measured value, and 3) the retrieved humidity had to be positive. The rms deviations of restored humidity values from measured profiles were approximately two times less than natural variations. A limited comparison with conventional linear statistical inversion showed that the nonlinear method may improve the recovery of vertical structure.
