Tailoring circulation type classification outcomes

• Published in: International journal of climatology 2021-11, Vol.41 (14), p.6145-6161
• Main Authors: Hansen, Felicitas, Belušić, Danijel
• Format: Article
• Language: English
• Subjects: atmosphere; Atmospheric circulation; Circulation; circulation patterns; Circulation types; Classification; clustering; Consistency; data partitioning; data standardization; Dependent variables; Euclidean geometry; Methods; Partitioning; Pressure distribution; Procedures; Sea level pressure; similarity measures; Spatial data; synoptic climatology; User requirements; weather types
• ISSN: 0899-8418; 1097-0088
• DOI: 10.1002/joc.7171

Efforts to intercompare many existing circulation type classification (CTC) methods have found no consistency in their outcomes. Therefore, when confronted with a task to classify atmospheric circulation types, it is difficult to find clear guidelines. This study explores the ways of increasing consistency between existing methods and obtaining physically meaningful and practically useful results. By applying a range of CTC methods to sea‐level pressure fields over a Scandinavian domain, it is shown that CTC methods using the same similarity measure (pattern correlation (CORR) or Euclidean distance (DIST)) have higher consistency. It is further shown that CTC outcomes can be tailored towards specific user requirements by properly manipulating the input data. Using unprocessed input data in DIST‐based CTC methods frequently results in classes containing physically inconsistent members because the classification procedure is obfuscated by circulation‐irrelevant information in the data. Using spatially standardized data in DIST‐based methods leads to considerably improved agreement with CORR‐based methods and brings high physical consistency within the individual classes. However, standardizing the input data removes too much of the circulation‐relevant information and results in no clear improvement in partitioning dependent variables such as precipitation. Best performance is achieved with DIST‐based methods using the input data with the spatial mean removed. This simple procedure focuses the CTC methods to use only the circulation‐relevant information and hence results both in physically consistent classes and in optimally performing partitioning of dependent variables. Consequently, the recommended guideline would be to use DIST‐based methods with spatial‐mean‐removed input data as the generally most effective classification approach. There is generally low consistency between the outcomes of different circulation type classification (CTC) methods. This study explores the ways of increasing consistency between existing methods and at the same time obtaining physically meaningful and practically useful results. We show that CTC outcomes can be tailored towards user requirements by properly manipulating the input data and recommend a simple procedure that results in physically consistent clusters as well as optimal partitioning performance for dependent variables.