Improving estimation of spatial precipitation in a mountain region

Abstract

Due to its importance for water resources, as well as flood and drought planning, an improved understanding of spatial precipitation patterns in mountain regions is needed. Precipitation gauge networks are sparse and traditional methods of interpolation yield inadequate precipitation fields for sparsely gauged mountain catchments. This work builds on a new method, Random Mixing, to generate multiple random spatial daily precipitation fields, conditioned on gauge observations. The Random Mixing algorithm has so far been tested on larger, densely gauged catchments. This project adapts the approach for a sparsely gauged, small 9.4 km2 mountain catchment, Marmot Creek Research Basin (MCRB) in Alberta, Canada. Three modifications have been made to the Random Mixing method in developing the new technique, which is referred to as RM-mountain: (1) improving spatial covariance, (2) introducing elevation dependence and (3) evaluating seasonal effects. Addition of each modification in turn increases the spatial variance of precipitation values across simulated fields. Leave-one-out cross-validation was used, and results compared with outputs from four deterministic spatial interpolation techniques. The best fit precipitation time series simulated by the RM-mountain generated ensemble members demonstrated improved precipitation estimates compared to the four deterministic techniques. Precipitation totals across the MCRB catchment generated by RM-mountain are higher than those from the other methods tested. Due to its random nature, RM-mountain enables generation of precipitation within the catchment on days when the gauges are dry. In contrast, deterministic spatial interpolation methods yield zero precipitation across the entire catchment on days with zero observed precipitation. Inclusion of modifications 1-3 in RM-mountain noticeably increased the likelihood of simulating more realistic precipitation values within the generated ensemble. To optimise selection of the most plausible fields, ensemble hydrological simulations were run, using a modified spatially-distributed version of the HBV conceptual model, and the physically-based Cold Regions Hydrological Model (CRHM), with a 200-member ensemble of time series spatial precipitation fields generated on a 50 m x 50 m regular model grid. Optimisation involved the use of Nash-Sutcliffe Efficiency (NSE) and bias metrics, to identify a best constructed time series that most closely simulates the observed streamflows. The improvement in streamflow bias with HBV was from -20.94 to 0.14; with CRHM, bias was 2 improved slightly from 2.04 to 1.88. Increases in NSE values were from 0.76 to 0.96 with HBV and from 0.54 to 0.74 with CRHM. Some noticeable differences between catchment responses with HBV and CRHM were observed, relating to the complexity of the models, i.e., the relative simplicity of the conceptual HBV model in contrast to the more complex physically-based CRHM. Notable examples of these differences were snowmelt earlier in the year and much less variation in the streamflow ensemble with HBV. A much greater variety of streamflow hydrographs in the CRHM-generated ensemble were due to CRHM’s much higher sensitivity to differences in observed meteorological input data, particularly wind speed. This work demonstrates that modifying a random method, by adapting how it randomly samples from observed precipitation at a small number of gauges, includes elevation gradients and seasonal variation, improves estimation of spatiotemporal precipitation patterns for a small mountain catchment and improves hydrological simulations. The new method has the potential to be used to enhance precipitation datasets to improve water resource and flood modelling in other sparsely gauged mountain regions.