A Neural Network to Retrieve the Mesoscale Instantaneous Latent Heat Flux over Oceans from SSM/I Observations

A new method is proposed to infer the instantaneous latent heat flux from satellite data. It is based on a neural network approach. The selection of the input parameters is physically based, and the sensitivity of the network to the latent heat flux is explained, by linearizing the microwave radiative transfer equation. The network is compared to the Liu and Niiler [1984] method on a global dataset grouping twelve ECMWF operational analyses and SSM/I brightness temperature measurements. The retrieval error found is thirty five W.m-2 for the network, which is fifteen W.m-2 less than with the Liu and Niiler [1984] method. For both algorithms, a spatial and zonal analysis shows that the variability of the retrieval error is less important in the tropical and equatorial regions than in sub tropical areas. An analysis of the error as a function of the wind speed and the vertical humidity gradient is also carried out, showing that the accuracy of the retrievals is better when the atmospheric humidity content is smaller. In the last part, the two methods are validated using data from the SEMAPHORE [Eymard et al., 1996], CATCH/FASTEX [Eymard et al., 1999], and TOGA/COARE [Webster and Lukas, 1992]. The retrieval error found on each validation dataset is consistent with the results obtained on the global dataset.