Comparison of snow distribution methods in the Echaurren Basin, Chilean Andes

Abstract

Neural networks, binary regression trees, and simple interpolation methods were employed to create snow water equivalence (SWE) maps for the 4.7 km2 Echaurren basin in the Chilean Andes (3300m elevation, 33.58° S, 70.13° N). Distributed SWE is critical for forecasting seasonal runoff and provides the initial condition for forecasting the timing of runoff. Data from five annual peak-accumulation snow surveys (1992-1996) involving snow depth measurements at approximately 100 points and representative density measurements were used to estimate SWE at each point in the basin (5-m grid spacing). Independent variables in the regression tree and neural network were elevation, slope, aspect, mean daily radiation, and soil type. Results developed using a regression tree approach were found to be very sensitive to the accurate location of survey points. A shift of as small as 10 m in the placement of survey points in the regression calculation gave a considerably different distributed SWE map for the basin. Both regression trees and neural networks produced qualitatively similar distributions of snow. Unlike neural networks, however, SWE maps from regression trees are limited to the range of input SWE values from field survey data. The neural network, on the other hand, can extrapolate SWE values in the basin. This is important in steeper slopes where the regression trees overestimated SWE. A comparison of errors using synthetic data for the catchment suggests that the neural networks gives a more accurate estimation of total SWE and distributed SWE for this catchment. Thiessen polygons showed similar SWE distributions to the regression tree and neural network distributed SWE maps. Due to the limited data points, polygons were constructed over extreme elevation ranges and may not accurately represent true basin SWE accumulation. Point and distributed energybalance models, in addition to a degree-day approach were used to simulate the melting of several distributed SWE maps. The snowmelt analysis was conducted on an hourly time step from peak-accumulation (September 30, 1992) until the end of the melt season (April 30, 1992). Unrouted meltwater generated using these models was then used to create simple hydrographs for the watershed, which were then compared to the observed basin hydrograph. Results of this comparison indicate that the distributed SWE map generated with the neural network method and melted with the distributed energy balance model most accurately matches the timing of observed runoff. The model needs to be adjusted in order to forecast more accurately the volume of Echaurren basin runoff.