Comparison of Regression Algorithms for the Retrieval of the Wet Tropospheric Path

This paper addresses the subject of the regression models used for the wet troposphere path delay correction for range measurements by satellite radar altimeters. The objective of this study is twofold: 1) to find which regression method is better suited for the retrieval between a neural network algorithm and a log-linear regression and 2) to determine whether the use of the altimeter backscattering coefficient at Ka- or Ku-band can substitute for the use of the radiometer brightness temperature at 18 GHz as an input for the retrieval. Several configurations of algorithms, including those used in the operational processing of altimetry missions such as Jason-1 or Envisat, are built and compared on the same learning and test database to determine which retrieval strategy is more appropriate. The importance of each input is analyzed and the performances of the different algorithms are assessed in terms of bias and standard of the errors and also in terms of their geographical distribution and correlations with other environmental variables. Their performances are then assessed on Jason-2 radiometer measurements using the criterion of variance in sea-surface height differences at crossovers. The study shows that the neural network formalism is better suited for the retrieval of the wet tropospheric path delay than the log-linear regression. In terms of variable selection, better results were obtained when the brightness temperature at 18 GHz was used instead of the backscattering coefficient. Overall, the best results are obtained with the combination of a three-channel radiometer and a neural network algorithm.