Developing the Snow Cover Fraction Schemes for Land Surface Models Using a Machine Learning Approach

Reconstruction of snow cover fraction (SCF) using historical records such as snow depth (SD), snow water equivalent (SWE), is important for investigating long-term snow cover variations under climate change. The SCF-SD scheme, when used in climate models, substantially affects the snow albedo feedback strength, the energy balance on ground bare soil and vegetation canopy thereby determining the simulation results for snow water equivalent (SWE). The SCF-SD relationship is strongly affected by surface roughness and subgrid topography and varied with snow accumulation/ablation seasons. However, the SCF schemes are not adequately developed, assessed for lack of accurate, high-resolution data. In addition, it is still unclear regarding whether other hydrologic, atmospheric or physiographic factors can be informative compared to SD/SWE to be incorporated as dependent variables in SCF formulation.

We apply the University of Arizona (UA) ground-based daily 4-km SD/SWE dataset, the Moderate Resolution Imaging Spectroradiometer (MODIS) regular latitude-longitude global grid daily SCF dataset (Terra, collection 6 version) and the Interactive Multi-sensor Snow and Ice Mapping System (IMS) gridded daily 1-km snow cover extent (SCE) dataset to develop the SCF formulations over 4-km4-km spatial resolution based on two existing schemes for land surface modeling. We adopt traditional regression approach and machine learning algorithms to incorporate more information (variables) suggesting new SCF schemes based on the computed mutual information between SCF estimates provided by a proposed model/hypothesis and the observed data. Prior to coupling the Noah-MP with WRF-Hydro for testing how different SCF formulations affect the streamflow hydrograph over a snow-dominated catchment, we evaluate all formulations in offline Noah-MP configuration by examining 1) the seasonal, interannual, and interbasin variations in SCF, SD and SWE against reference datasets, 2) the simulated long-term trend of SD and SWE against UA snow product, and 3) the changes of net radiation, turbulent fluxes, ground heat fluxes and soil/ground temperature for accumulation and ablation seasons.