DETERMINATION OF CLOUD LIQUID WATER PATH OVER THE OCEANS FROM SPECIALSENSOR MICROWAVE IMAGER (SSM/I) DATA USING NEURAL NETWORKS/

A neural network (NN) has been developed in order to retrieve the clou d liquid water path (LWP) over the oceans from Special Sensor Microwav e/Imager (SSM/I) data. The retrieval with NNs depends crucially on the SSM/I channels used as input and the number of hidden neurons-that is , the NN architecture. Three different combinations of the seven SSM/I channels have been tested. For all three methods an NN with five hidd en neurons yields the best results. The NN-based LWP algorithms for SS M/I observations are intercompared with a standard regression algorith m. The calibration and validation of the retrieval algorithms are base d on 2060 radiosonde observations over the global ocean. For each radi osonde profile the LWP is parameterized and the brightness temperature s (Tb's) are simulated using a radiative transfer model. The best LWP algorithm (all SSM/I channels except T85V) shows a theoretical error o f 0.009 kg m(-2) for LWPs up to 2.8 kg m(-2) and theoretical ''clear-s ky noise'' (0.002 kg m(-2)), which has been reduced relative to the re gression algorithm (0.031 kg m(-2)). Additionally, this new algorithm avoids the estimate of negative LWPs. An indirect validation and inter comparison is presented that is based upon SSM/I measurements (F-IO) u nder clear-sky conditions, classified with independent IR-Meteosat dat a. The NN-based algorithms outperform the regression algorithm. The be st LWP algorithm shows a clear-sky standard deviation of 0.006 kg m(-2 ), a bias of 0.001 kg m(-2), nonnegative LWPs, and no correlation with total precipitable water. The estimated accuracy for SSM/I observatio ns and two of the proposed new LWP algorithms is 0.023 kg m(-2) for LW P less than or equal to 0.5 kg m(-2).