A methodology for generating and validating downscaled estimates of meteorological data at the basin scale
Abstract
Meteorological data (e.g., precipitation and temperature) serve as critical forcing inputs for hydrological models. Such data are currently available at ~1km resolution (i.e., PRISM and Daymet) for the continental United States, whereas higher resolution inputs are needed for enhanced predictive capabilities. Here, we present a machine learning framework to generate and validate downscaled estimates of precipitation and temperature in the East Taylor subbasin (in Colorado Rockies, United States). We use Random Forests to extract the relationships between high-resolution topographic variables and precipitation/temperature at ~1 km resolution, and leverage these relationships to generate downscaled estimates at resolutions of 400m, 200m, and 100m. We regularize our model by using nearest-neighbor maps developed from point measurements. Subsequently, we validate the downscaled estimates using a two-pronged approach. First, we investigate if the downscaled estimates of precipitation/temperature exhibit lower values of squared-error when compared to point measurements. Second, we develop a data-driven approach to model spatially distributed snowpack (obtained via the Airborne Snow Observatory) using spatially distributed precipitation/temperature as input. We used this approach to investigate if using downscaled estimates of precipitation/temperature can help model spatially smooth estimates of snowpack that are also more accurate when compared to using the original ~1km resolution estimates. We observed that our downscaled estimates were the most reliable at 400m resolution, with reliability reducing as we go to finer scales. We conclude that while topography-based variables are adequate for downscaling precipitation and temperature down to 400m resolution, additional high-resolution variables (such as surface reflectance, land use, and vegetation) are needed to reach finer scales. This work was supported by the DOE Office of Science, ExaSheds Project.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H34F..08M