Statistical atmospheric downscaling of short-term extreme rainfall by neural networks

Abstract

Statistical atmospheric rainfall downscaling, that is, statistical estimation of local or regional rainfall on the basis of large-scale atmospheric circulation, has been advocated to make the output from global and regional climate models more accurate for a particular location or basin. Neural networks (NNs) have been used for such downscaling, but their application has proved problematic, mainly due to the numerous zero-values present in short-term rainfall time series. In the present study, using serially coupled NNs was tested as a way to improve performance. Mean 12-hour rainfall in the Chikugo River basin, Kyushu Island, Southern Japan, was downscaled from observations of precipitable water and zonal and meridional wind speed at 850 hPa, averaged over areas within which the temporal variation was found to be significantly correlated with basin rainfall. Basin rainfall was ranked into four categories: no-rain (0) and low (1), high (2) and extreme (3) intensity. A series of NN experiments showed that the best overall performance in terms of hit rates was achieved by a two-stage approach in which a first NN distinguished between no-rain (0) and rain (1–3), and a second NN distinguished between low, high, and extreme rainfalls. Using either a single NN to distinguish between all four categories or three NNs to successively detect extreme values proved inferior. The results demonstrate the need for an elaborate configuration when using NNs for short-term downscaling, and the importance of including physical considerations in the NN application.