Improving Monsoon Precipitation Forecast Accuracy over Arizona with GPS-PWV Data Assimilation

Forecasting monsoon precipitation over Arizona is challenging, partly due to its complex terrain. The model grid structure may misrepresent topographic details, and the sparse observation network is insufficient for the initialization of the model at the scale of the topography (;4 km), particularly in defining the spatial distribution of moisture. Our study aims to assess monsoon precipitation forecast skill over Arizona by simulating 24 precipitation events of the 2021 monsoon season in convective-permitting Weather Research and Forecasting (WRF) Model simulations from a 40-member ensemble coupled with data assimilation (DA). The High-Resolution Rapid Refresh (HRRR) model is used as the initial and boundary conditions. The data being assimilated are hourly global positioning system precipitable water vapor (GPS-PWV) data, collected from 31 sites over the Southwest United States, including special observations collected in Arizona during the 2021 monsoon. The configuration of the simulations is based on seven experiments in which microphysics schemes and horizontal localizations were varied. Our results show: 1) GPS-PWV data assimilation reduced forecast PWV errors and biases during the DA cycle and forecast periods; 2) the assimilation increased the instability of the preconvective atmosphere due to moistening over the Mogollon Rim and southeastern Arizona by as much as 1000 J kg²¹, persisting for at least 6 h into the forecasts; and 3) the assimilation improves the precipitation forecast skill up to 9 h into the forecasts and lowers the biases in the temperature, dewpoint, and mixing ratio within at least 3 km above ground level. SIGNIFICANCE STATEMENT: Forecasting rainfall is difficult due to complex land–atmosphere interactions. Our study assesses the skill of a high-resolution numerical weather prediction (NWP) model in forecasting monsoon rainfall over Arizona. By better specification of the initial atmospheric moisture through assimilating an observed moisture dataset from global positioning system (GPS) receivers, our study finds that the errors and biases in the modeled moisture are reduced at the initial hour of the forecasts, the modeled atmospheric instability increases, i.e., it is more favorable for convection to occur during the first 6 h of the forecasts, and the modeled precipitation is more comparable with the observed precipitation in the first 9 h into the forecasts.