Improving the quality of INSAT derived quantitative precipitation estimates using an neural network method

Abstract

 In this paper an Artificial Neural Network (NN) approach has been applied to improve the quality of the INSAT derived sub-division quantitative precipitation estimates (IMD-QPE) over the Indian region for the summer monsoon season. Data for the years 2001, 2003 and 2004 have been used as the training sample. The method is tested with independent sample data for the year 2005. For the subdivisions over the domains of high orographic and monsoon low pressure system, where very rainfall occasionally occurs, different network architectures are applied to minimize the IMD-QPE errors. An inter-comparison between NNQPE (NN model output IMD-QPE), IMD-QPE and actual rainfall indicates that the pattern of NNQPE is closer to the observed rainfall distribution. The weekly mean absolute error of IMD-QPE with respect to observed rainfall, which ranges between 10–99 mm, becomes 4–70  mm in case of NNQPE. The performance statistics shows that the proposed NN model is able to produce better IMD-QPE with higher skill score and correlation co-efficient with respect to observation in most of the sub-divisions. The method is found to be promising for operational application.