Development and evaluation of new weather types classification for Croatia

Weather type classification in Croatia has been based on subjective Poje’s classification since 1965. In the first stage of the analysis the results of the COST733 action “Harmonization and Applications of Weather Types Classifications for European Regions” were applied to the Croatian area to improve previous knowledge and practice, and to make the whole classification process automatized. Based on the obtained results and conclusions from the first stage, analysis in the second stage is done by running COST733class software for different settings. This analysis aims to determine the most suitable combination of main classification parameters such as domain size, number of types, input variables and classification methods to develop a classification that is intended to be the most appropriate to capture precipitation conditions in Croatian regions. The assessment of evaluation results takes into account the effect of varying numbers of types and specifies temporal (among seasons) and as well spatial (among regions) variations. For this purpose, Croatia is divided into five regions to acknowledge the orography and climate diversity of the Croatian area which lead to different weather conditions within the same atmospheric circulation. Two regions belong to the continental part of Croatia (Eastern and Central Croatia), while the rest, including mountainous parts, are placed along with the Adriatic sea (North, Middle and South Adriatic). The efficacy of different objective classifications developed using COST733class is examined utilizing Croatian meteorological station data. For evaluating the discriminative power (synoptic skill) of the classifications and the relevance of specific settings several statistical metrics are used and the statistical significance of the differences in statistic metrics between classifications is tested. The results from the first stage have shown better classification performance for optimization and threshold based methods among other methods, hence we proceed with GWT (GrossWetterTypes) and DKM (dkmeans) methods in the second stage. Results for all different settings have shown better classification performance along the Adriatic coast and in the mountainous parts than in the more continental parts. Further, some preliminary results show that classifications based on DKM methods perform better when using input data with spatial-mean-removed and that domain size and position should be tailored for each input variable. Also, the improvement of classification is found when including some of the additional variables together with the main variable (usually mean sea level pressure). However, it is not clear the inclusion of which additional variable is the most relevant and statistically significant, and how many of them should be included at the same time.