Intercomparison of inversion algorithms to retrieve rain rate from SSM/I by using an extended validation set over the Mediterranean area

The capability of some inversion algorithms to estimate surface rain rate at the midlatitude basin scale from the Special Sensor Microwave Imager (SSM/I) data is analyzed. For this purpose, an extended database has been derived from coincident SSM/I images and half-hourly rain rate data obtained from a rain gauge network, placed along the Tiber River basin in Central Italy, during nine years (from 1992 to 2000). The database has been divided in a training set, to calibrate the empirical algorithms, and in a validation one, to compare the results of the considered techniques. The proposed retrieval methods are based on both empirical and physical approaches. Among the empirical methods, a regression, an artificial feedforward neural network, and a Bayesian maximum a posteriori (MAP) inversion have been considered. Three algorithms available in the literature are also included as benchmarks. As physical algorithms, the MAP method and the minimum mean square estimator have been used. Moreover, in order to test the behavior of the algorithms with different kinds of precipitation, a classification of rainy events, based on some statistical parameters derived from rain gauge measurements, has been performed. From this classification, an attempt to identify the type of event from radiometric data has been carried out. The purposes of this paper are to determine whether the use of an extended training set, referred to a limited geographical area, can improve the SSM/I skill in rain detection and estimation and, mainly, to confirm the validity of the physical approach adopted in previous works. It will be shown that, among all the estimators, the neural network presents the best performances and that the physical techniques provide results only slightly worse than those given by empirical methods, but with the well-known advantage of an easy application to different geographical zones and different sensors