The use of similarity concepts to represent subgrid variability in land surface models: Case study in a snowmelt-dominated watershed

This paper develops a multivariate mosaic subgrid approach to represent subgrid variability in land surface models (LSMs). The k-means clustering is used to take an arbitrary number of input descriptors and objectively determine areas of similarity within a catchment or mesoscale model grid box. Two different classifications of hydrologic similarity are compared: an a priori classification, where clusters are based solely on known physiographic information, and an a posteriori classification, where clusters are defined based on high-resolution LSM simulations. Simulations from these clustering approaches are compared to high-resolution gridded simulations, as well as to three common mosaic approaches used in LSMs: the "lumped" approach (no subgrid variability), disaggregation by elevation bands, and disaggregation by vegetation types in two subcatchments. All watershed disaggregation methods are incorporated in the Noah Multi-Physics (Noah-MP) LSM and applied to snowmelt-dominated subcatchments within the Reynolds Creek watershed in Idaho. Results demonstrate that the a priori clustering method is able to capture the aggregate impact of finescale spatial variability with O(10) simulation points, which is practical for implementation into an LSM scheme for coupled predictions on continental-global scales. The multivariate a priori approach better represents snow cover and depth variability than the univariate mosaic approaches, critical in snowmelt-dominated areas. Catchment-averaged energy fluxes are generally within 10%-15% for the high-resolution and a priori simulations, while displaying more subgrid variability than the univariate mosaic methods. Examination of observed and simulated streamflow time series shows that the a priori method generally reproduces hydrograph characteristics better than the simple disaggregation approaches.