Clustering and upscaling of station precipitation records to regional patterns using self-organizing maps (SOMs)

Abstract

Self-organizing maps (SOMs), a particular application of artificial neural networks, are used to proportionately combine precipitation records of individual stations into a regional data set by extracting the common regional variability from the locally forced variability at each station. The methodology is applied to a 100 yr record of precipitation data for 104 stations in the Mid-Atlantic/Northeast United States region. The SOM combines stations with common precipitation characteristics and identifies precipitation regions that are consistent across a range of spatial scales. A variation of the SOM application identifies the temporal modes of the regional precipitation record and uses them to fill missing data in the station observations to produce a regional precipitation record. A test of the methodology with a complete data set shows that the 'missing data' routine improves the regional signal when up to 80% of the data are missing from 80% of the stations. The improvement is almost as pronounced when there is a bias in the missing data for both high-precipitation and low-precipitation events.