A real-time snow water equivalent interpolation system using wireless sensor networks and historical remotely sensed data

Abstract:

Over 100 wireless sensors for monitoring real-time snow conditions were deployed in ten clusters distributed across the headwaters of the American River Basin. The sensors are strategically placed to measure snow depth across elevation gradients and local differences in slope, aspect and canopy coverage. The sensors provided near-real-time snow-depth readings during the 2014 and 2015 snow seasons. Also, time series of snow water equivalent (SWE) maps were reconstructed for 2012-2014 using energy-balance modeling with modeled energy forcings (NLDAS), terrain corrections for solar radiation (TOPORAD), and fractional snow cover data (MODIS). We blended the real-time snow-depth readings with the historical SWE reconstructions to interpolate real-time SWE conditions across the basin. Snow-depth readings from all sensors for selected dates in 2014 were converted into SWE estimates using density values from snow-pillow sites in the basin. SWE values for pixels where sensors were located were extracted from the reconstructed 2012-2013 data to develop a time-series array. Using a Nearest-Neighbor algorithm we searched the array for the closest conditions that matched the sensor data, interpolated the residuals between reconstructed versus measured SWE across the basin, and added the interpolated values to the reconstructed SWE. We also blended the sensor measurements with the 2014 reconstruction results that were from the same dates. We evaluated both the historical SWE blending results and the concurrent SWE blending results with the operational networks measurements, finding that the concurrent SWE blending has a slightly lower RMSE compared to that of historical SWE blending. Since the reconstruction results could only be estimated after the end of the season, concurrent SWE blending is not applicable to the real time SWE interpolation even though it has a more accurate estimation. However, the small difference of RMSE between the two approaches informs us that the historical blending is better fitted for real time SWE interpolation.