Generating high-resolution estimates of precipitation at the watershed scale using machine learning

Abstract:

Precipitation data serves as a critical forcing input for ecohydrological models. Precipitation data is currently available at 800 m resolution (PRISM) for the continental United States, whereas high-resolution precipitation (<200 m) inputs are needed for enhanced predictive capabilities. Here we present a machine learning framework to generate high-resolution estimates of precipitation by downscaling coarse-resolution (~12.5 km NLDAS) data in the Upper Colorado Water Resource Region (UCWRR). Our approach uses Random Forests and Convolutional Neural Networks and has four steps. First, we impute or “gap-fill” incomplete records from ground weather stations (e.g., from NOAA and NRCS networks). We do this using a newly developed sequential imputation algorithm that leverages other incomplete records to impute missing data. Second, we downscale the coarse-resolution precipitation to an intermediate (~800 m) resolution using existing products to establish a cross-scale relationship. We do this by incorporating the effects of various factors such as elevation, aspect and land-use. Third, we downscale precipitation from the intermediate resolution to high-resolution using the imputed or measured weather station data. Finally, we validate the downscaled estimates using high-resolution meteorological data products (e.g., NASA ASO data) and cross-validation. Preliminary results indicate that the proposed approach is capable of generating high-resolution precipitation data with reasonable accuracy. This work was supported by the DOE Office of Science, ExaSheds Project.