Objective determination of the presence of lake-effect zones in climate models

In this study, we use K-means clustering to develop a methodology to determine the presence of lake-effect precipitation zones in model simulations of the Laurentian Great Lakes Basin. Our goal is to have a quick evaluation method that can identify if the qualitative representation of precipitation is improving as modeling technology advances. We use two regional climate model ensembles (North American Coordinated Regional Downscaling Experiment and the International Centre for Theoretical Physics’ Regional Climate Model 4 as coupled to a 1D lake model with some modified at the University of Wisconsin-Madison). University of Delaware (UDel) observational monthly winter precipitation data from the basin are compared to the model ensemble simulations over the climatology period of 1980–1999 using an object-based K-means cluster analysis and performance tools from the METplus evaluation package. Lake-effect zones are present by visual inspection in the Udel observations, and a choice of four clusters represents lake-effect zones in the observations across the basin. For the model simulations, it is not possible at the basin scale to identify the lake-effect zones visually. The application of the K-means clusters at the basin scales extracts geographical features in the correct locations; however, the features are biased compared with observations. The analysis is repeated at the sub-basin scale, and the lake-effect zones are revealed well enough to provide insights into model performance. The high sensitivity of the K-means cluster analysis to the domain size, regional, basin-scale, and sub-basin scale, is greater than expected. However, it reveals important aspects of the performance of present-generation simulations. The paper discusses the K-means cluster analysis and its performance metrics provided by METplus, systematic simulation bias at the basin and regional scale, and representation of lake-effect zones by model ensemble members at the sub-basin scale. We conclude that the clustering techniques could quicken the identification of lake-effect zones; however, significant reduction of simulation bias is needed to improve the representation of lake-effect zones prior to detailed scientific investigation and application to regional climate adaptation applications.