Improvement of Summer Precipitation Simulation Over the Tibetan Plateau: Implications for Assimilating Pseudo Soundings Generated by Machine Learning

Abstract Due to the sparse observations over the Tibetan Plateau (TP), initial conditions in atmospheric models over the TP have large uncertainty, which could degrade the simulation performance of precipitation over the TP. This study proposes a novel approach for generating “pseudo soundings” for the atmospheric model's initial conditions. In each “pseudo soundings”, the vertical profiles of heat and moisture are estimated by a random forest algorithm from the evolution and mean state of precursor convective precipitation. Results indicate that the vertical structures of temperature and humidity in “pseudo soundings” estimated in this study could be superior to the reanalysis data set, and could well remedy the insufficient heat and moisture information for model's initial condition over the TP where in situ conventional soundings are absent. Compared to the ERA5 reanalysis data set, the mean relative error of temperature and specific humidity estimated in this study could be reduced by 0.7% and 5.7%, respectively. Verification based on a series of simulations illustrates that assimilating vertical profiles of heat and moisture in “pseudo soundings” would alleviate the spurious and overestimated summer precipitation in simulations over the TP. The mean simulation bias of hourly precipitation could be reduced by 31.8%, compared to assimilating in situ conventional soundings only. Further investigations suggest that assimilating vertical profiles of heat and moisture in “pseudo soundings” would improve the simulation performance of summer precipitation by producing a dryer and more convective stable lower atmosphere over the TP.