A physically-based multivariate-regression approach for downscaling NEXRAD precipitation in mountainous terrain

Precipitation temporal and spatial variability often controls terrestrial hydrologic processes and states. Common remotely-sensed precipitation products have a spatial resolution that is often too coarse to reveal hydrologically important spatial variability. A parsimonious physically-based multivariate-regression algorithm, referred to as multi-level cluster-optimizing ASOADeK regression, coupled with a random cascade model, is developed for downscaling low-resolution spatial precipitation fields. This algorithm auto-searches precipitation spatial structures (e.g., rain cells), and atmospheric and orographic effects, to estimate precipitation distribution without prior knowledge of the atmospheric setting. The only required input data for the downscaling algorithm are a large-pixel precipitation map and the DEM map of the area of interest. We tested the algorithm by downscaling NEXRAD-derived 16 km x 16 km precipitation fields to 4 km * 4 km pixel precipitation for both daily and hourly precipitation in the northern New Mexico mountainous terrain and the central Texas hill country. The algorithm generated downscaled precipitation fields in good agreement with the original 4 km * 4 km NEXRAD precipitation field, as measured by precipitation spatial structures and the statistics between downscaling and the original NEXRAD precipitation maps.