Temporally downscaling a precipitation intensity factor for soil erosion modeling using the NOAA-ASOS weather station network

Precipitation intensity is an important meteorological input for water erosion and runoff applications. A commonly used intensity factor is maximum 30-min intensity (I-30), which represents the sustained intensity of a storm. Determining I-30 is challenging for two reasons. First, intensity can vary significantly over time, even within very short durations of 5 min or less. Second, the majority of precipitation data sets are limited by their fixed-interval nature, and I-30 may not be constant within fixed measurement intervals. When intensity is simply averaged given the accumulation of a measurement interval, the temporal resolution of the precipitation data set biases the result. Therefore, in this study, bias adjustments were determined for a range of selected temporal resolutions and Koppen-Geiger climate regions in the United States. In this case, the intensity factor was monthly mean maximum 30-minute intensity (MX.5P), which is a parameter used to generate stochastic meteorological inputs for models that include the Rangeland Hydrology and Erosion model (RHEM) and the Water Erosion Prediction Project model (WEPP). The adjustment factors were obtained by using linear regression of reference MX.5P values derived from breakpoint data against MX.5P values aggregated from the breakpoint data to represent lower temporal resolutions. The resulting slope coefficients were used to determine bias adjustment factors. In addition, multivariate machine learning regression was used to obtain more complex correlations involving a host of predictor variables that may each be determined from daily precipitation statistics and the spatial location of each station. In total, 609 stations and 16 climate classifications were represented in the regressions. Linear regressions for climate classifications gave RMSE for values of MX.5P derived from hourly data ranging from 0.98 to 3.46 mm hr(-1) with an average of 2.18 mm hr(-1). For daily data, the error range was 2.83-8.44 mm hr(-1) with an average of 5.61 mm hr(-1). The multivariate regression using machine learning algorithms improved regressions for coarser resolutions, reducing error to the 3-4 mm hr(-1) range for downscaled daily values.