Rainfall Nowcasting Algorithm Based on Radar and Satellite Data

A real time algorithm for predicting the spatial distribution of rainfall rate is introduced in this work. The suggested algorithm uses radar reflectivity and satellite brightness temperature data to predict the rainfall field. The introduced method forecasts first the most likely rain areas and then predicts the expected intensity of rainfall rate in each pixel. The algorithm searches for contiguous rain pixels and tracks the main rain cells. Assuming first that in a short time (15 minutes) a rain cloud approximately behaves as a rigid object, all parts of the cloud move in the same direction and at a constant speed. Based on this assumption the cloud motion vector of a rain cell can be estimated by using centroid displacement of a single cell observed in two consecutive radar images. The motion vectors are used to advect the rain cells and determine the potential rainfall areas. Assuming also that a current reflectivity is related with previous radar and satellite images located in a neighborhood region with center on a predicted pixel, an exponential regression model is proposed to predict the growth and decay of rainfall intensity. The predicted rainfall areas are segregated into small regions to best represent the nonstationary and stochastic rainfall variability. The regression model uses a spatial and temporal exploration and includes a large number of predictors. Thus, an iterative forward selection algorithm is introduced to eliminate irrelevant pixels and determine the best predictors for each region. The algorithm makes predictions at every 15 minutes and lead time varies from 15 to 60 minutes. Five tropical storms that occurred in Puerto Rico during 2003-2013 were used to validate the proposed algorithm. Results show the proposed algorithm is a potential tool to couple with a hydrological numerical model to predict flash flood events.