A Comparison of TRMM and WRF model hindcasts of rainfall in the Caribbean

Extreme rainfall events have the capacity to cause significant loss of life and damage to infrastructure. From an insurance perspective, determining the frequency of such events, as well as assessing the return period of any specific event for the purpose of triggering a payout, are of paramount importance. Unfortunately, such determinations are difficult. They require long observation periods, preferably in excess of 30 years. The fragmented and regional nature of rain events is another difficulty. A further complication is the nature of rain observations. Although the rain gauge is considered the gold standard for rain observations, gauges are usually relatively far apart, especially in the developing world. Convective rain events are often quite spotty in nature—very often some gauges will show no precipitation while nearby gauges show extreme amounts, depending on the exact motion of a given storm cell. Finally, gauges are sometimes not checked or read consistently, have maintenance issues, or have been relocated over time, complicating the spatial and temporal analyses.

In support of the development of an extreme rainfall index system for the Caribbean Catastrophe Risk Insurance Facility, a comparison was made between remotely sensed data, numerical models, and rain gauges using data in the Caribbean and Central America, including Florida and Puerto Rico. This paper describes an analysis of the performance of the Tropical Rainfall Measurement Mission 3B42 data sets and various configurations of two dynamical cores in the Weather Research and Forecasting (WRF) model Version 3.3.1. After initial experiments, configurations were selected for the Advanced Research core (ARW) and the Nonhydrostatic Mesoscale Model (NMM) core. Various input boundary data sets were also run and the differences evaluated, focusing on real time GFS initial time, the GFS final analysis data sets, and data from the NCAR/NCEP Reanalysis Project.

For the final analysis configuration, a set of scripts was developed to automate the process of running single day hindcasts from the models for a given day, fetching the TRMM and observations, and generating a set of standard outputs designed to facilitate the assessment of the simulations. The standardized outputs consist of the modeled rainfall from each of the two WRF cores, the TRMM 3B42 rainfall, and available station reports for that day derived from the National Climate Data Center (NCDC) Global Summary of the Day (GSOD) archives. Individual station data were compiled and analyzed using the R statistics package. Correlations were computed between several output variables (Temperature, Dew Point, Mean Sea Level Pressure, and Rainfall), and a scatter plot then created and incorporated in a Google Earth file (KML) as well as traditional data sets for further analysis using JMP software. The KML format and Google Earth facilitated the quick review of any of the hundreds of simulations generated by this study.

Neither the TRMM data or either numerical model estimates generated what could be considered as good correlations with the individual daily gauge reports, especially for Caribbean stations. Correlations (Pearson's ρ) between gauge and either satellite or model results were on average below 0.5. However, this is not entirely surprising. Previous studies (Chokngamwong and Chiu, 2007, for example) show relatively poor correlations between rain gauges and modeled or satellite remote sensing precipitation estimates for specific events, with much better correlations for climatology or areal averages. This study verifies and expands these findings. There are several key reasons for this phenomenon, and the apparently poor direct correlation between the rain gauges and either modeling or remote sensing techniques must be carefully understood in the context of what exactly each sensor is measuring and the application to which it is to be applied. In particular, the fact that gauges are measurements at a point whereas both the satellite and model outputs are spatial averages is a significant source of the difference. In controlled analyses of closely located sets of stations (those within a single model or satellite grid cell), correlations rose dramatically. The study concluded that the hindcast produced by the NMM core produced the best correlations, closely followed by the ARW core. Both WRF configuration hindcasts were noticeably superior to the TRMM outputs when using GFS Final analysis data sets for boundary conditions. Comparisons with known extreme rainfall events in Guyana and Barbados were made, with the WRF hindcasts again proving superior to the TRMM data.