Retrieval of water vapour profiles from radio occultation refractivity using artificial neural network

Abstract

Artificial neural network (ANN) technique has been used to derive water vapour pressure profiles in the troposphere from radio occultation data over India and adjoining region. A fully connected three-layer network, with one hidden layer, has been constructed and standard back propagation algorithm has been used to train the network. While month, latitude and vertical profile of refractivity/bending angle constitute the input vector, the water vapour partial pressure profile forms the output vector. Only the moisture-laden summer monsoon months of June, July, August and September of 2010 have been considered for developing the retrieval algorithm. There are 2120 input and output pairs, out of which 1696 pairs form the training set while the remaining pairs constitute the validation set. The retrieved profiles of water vapour pressures in the validation set have been compared with the corresponding COSMIC operational products of water vapour pressure profiles. The effectiveness of the algorithm is apparent from this comparison and also from the vertical profiles of bias and root mean square error (RMSE). The statistics show better performance of the algorithm with refractivity as one of the inputs than with bending angle. The RMSE in water vapour retrieval from refractivity is within 1.5-2 hPa compared to markedly higher values of 6 hPa when derived from bending angle. Additionally, the algorithm has also been tested in an independent year 2009 and the performance of the refractivity based retrieval has been found to be highly consistent in the year 2010, with RMSE within 1.5 hPa.