On the Representativeness of Observational and Modeling Data in Numerical Weather Prediction

With increasing spatiotemporal weather observations (e.g., mesonet) and development of fine-scale numerical weather prediction, data assimilation and forecast verification that incorporates data errors at the ultra high-resolution raise questions about the representativeness of individual observations and gridded forecasts. Accurately estimating the representativeness can be very challenging, especially in mountainous areas. In this study, we have made use of massive observational meteorological data (intervals of 100s meters in space and 5 minutes in time) near Granite Peak, Utah collected by special projects in recent years, and corresponding weather forecasting data from an NCAR/RAL real-time weather forecasting system at ~100s meters resolution to investigate the data representativeness in space and time. To characterize the data representativeness, advanced statistics, such as vector correlation, varographs and integrated analytics like singular value decomposition, are used to study the probability distribution function (PDF) of these data. The adequate samples enable us to better estimate the data PDF and to better understand observational error, modeling error, and representativeness error. The talk will present the insights derived from estimating the numerical weather prediction data representativeness, which can be highly useful in forecast verification, data assimilation, station siting, and observational network distribution.