Official URL: https://doi.org/10.1007/s11269-020-02714-1

A challenge for climate impact studies is the identification of a sub-set of climate model projections from the many typically available. Sub-selection has potential benefits, including making large datasets more meaningful and uncovering underlying relationships. We examine the ability of seven sub-selection methods to capture low flow and drought characteristics simulated from a large ensemble of climate models for two catchments. Methods include Multi-Cluster Feature Selection (MCFS), Unsupervised Discriminative Features Selection (UDFS), Diversity-Induced Self-Representation (DISR), Laplacian score (L Score), Structure Preserving Unsupervised Feature Selection (SPUFS), Non-convex Regularized Self-Representation (NRSR) and Katsavounidis–Kuo–Zhang (KKZ). We find that sub-selection methods perform differently in capturing varying aspects of the parent ensemble, i.e. median, lower or upper bounds. They also vary in their effectiveness by catchment, flow metric and season, making it very difficult to identify a best sub-selection method for widespread application. Rather, researchers need to carefully judge sub-selection performance based on the aims of their study, the needs of adaptation decision making and flow metrics of interest, on a catchment by catchment basis