Assessing the history-based predictability of regional monthly precipitation data using statistical and fuzzy methods

Abstract

This paper points out two procedures for verifying the statistical independence between the historical and actual data series. They are based respectively on the computation of an empirical distance correlation coefficient, and the use of Hamming distance between binary encoded phase signals corresponding to Log-Gabor filtered time-series. Both procedures are applied to the monthly precipitation series recorded during 480 successive months at 49 meteorological stations in Dobrogea region (Romania) and on 4 and 5-valued fuzzified versions of the initial data represented using linguistic labels. The results show that the crisp data and their fuzzified versions cannot support the hypothesis of history-based predictability from their history. Hence, the two statistical independence tests are robust with respect to the k-means fuzzification of data and cross-validate each other, being applicable to any long-term analysis of precipitation series, to any other signals in general, and to their fuzzified versions, as well.