Pattern recognition of radar echoes for short-range rainfall forecast

A four-layer feed-forward back-propagation artificial neural network (ANN) is applied to weather radar echo maps of reflectivity data for the prediction of heavy rainfall events in the short-range of 1 to 2 hours. Inputs for the ANN are the cross correlations of statistical measures of a sequence of radar images. The ANN is trained to capture increasingly organized echo patterns that often are preludes to localized heavy rain. Results show that the ANN is able to achieve a success rate of 89% against a false alarm rate of 33%. In parallel, a separate module utilizing Hough transform is developed to depict the lining up of echoes on the reflectivity maps. The module provides an objective analysis tool for forecasters to test the hypothesis that crossing or merging of echo lines, the so-called “X” patterns, would lead to enhanced convection at preferred locations. Working in tandem, the ANN helps to isolate specific sectors on the radar maps where organization is taking place so that the Hough transform module (HTM) can be meaningfully applied in the appropriate target areas. In turn, parameters derived from the HTM, along with the standard statistical measures, can be fed back into the ANN for further training and system enhancement in the identification of “X” patterns