Classification of rainfall into small- and large-scale events by statistical pattern recognition

Abstract

Scale properties of daily areal rainfall in south-eastern Norway are investigated, and the events are classified into small- and large-scale events by statistical pattern recognition. The coefficient of variation and the standard deviation, computed from all precipitation gauges for one event, are considered to be mutually independent and are assigned gamma densities. The two statistical parameters form a feature vector which is used as the discriminatory variable in the classification algorithm. Empirical and theoretical semivariograms were fitted to the classified sets of events, and the parameters of the semivariogram (nugget, sill and range) were investigated for each set. Different values for all parameters were found, and the differences in range, describing the scale of the phenomena, were further investigated by the use of cross validation with the MSIE (mean squared interpolation error) criterion. Optimal range of daily small-scale precipitation was found to be 16 km (19 km without nugget effect), and 33 km (70 km without nugget effect) for daily large-scale precipitation. These optimal values for range, however, proved to be of small consequence for the estimated areal precipitation using kriging interpolation. Areal precipitation was computed for catchments of size 50 km 2 and 1200 km 2, and insignificant differences in the computed areal precipitation were found when applying the class-specific semivariograms instead of the mean semivariogram calculated from all events regardless of class. However, when areal reduction factors (ARFs) were derived from the analytical expressions of the class-specific semivariograms, considerable differences were found for the ARF curves of the two precipitation types. The difference became more marked for precipitation events of low probability of occurrence.