Verification of the skill of numerical weather prediction models in forecasting rainfall from U.S. landfalling tropical cyclones

Highlights

• The skill of NWP model in forecasting tropical cyclone (TC) rainfall is assessed.

• The focus is on 15 North Atlantic TCs.

• Five remote-sensing products are used to aid in the forecast evaluation.

• NWP models can skillfully forecast TC rainfall with lead-times up to 48 h.

Abstract

The goal of this study is the evaluation of the skill of five state-of-the-art numerical weather prediction (NWP) systems [European Centre for Medium-Range Weather Forecasts (ECMWF), UK Met Office (UKMO), National Centers for Environmental Prediction (NCEP), China Meteorological Administration (CMA), and Canadian Meteorological Center (CMC)] in forecasting rainfall from North Atlantic tropical cyclones (TCs). Analyses focus on 15 North Atlantic TCs that made landfall along the U.S. coast over the 2007–2012 period. As reference data we use gridded rainfall provided by the Climate Prediction Center (CPC). We consider forecast lead-times up to five days. To benchmark the skill of these models, we consider rainfall estimates from one radar-based (Stage IV) and four satellite-based [Tropical Rainfall Measuring Mission - Multi-satellite Precipitation Analysis (TMPA, both real-time and research version); Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN); the CPC MORPHing Technique (CMORPH)] rainfall products. Daily and storm total rainfall fields from each of these remote sensing products are compared to the reference data to obtain information about the range of errors we can expect from “observational data.” The skill of the NWP models is quantified: (1) by visual examination of the distribution of the errors in storm total rainfall for the different lead-times, and numerical examination of the first three moments of the error distribution; (2) relative to climatology at the daily scale. Considering these skill metrics, we conclude that the NWP models can provide skillful forecasts of TC rainfall with lead-times up to 48 h, without a consistently best or worst NWP model.

Introduction

North Atlantic tropical cyclones (TCs) are responsible for significant societal and economic impacts. Over the 1900–2005 period, the average annual normalized damage associated with TCs in the continental United States is about $10 billion (value normalized to 2005 monetary value; Pielke et al., 2008). Overall, this damage accounts for close to half (Table 1 in Smith and Katz (2013)) of the total weather and climate disasters over the period of 1981–2011, much more than the damage associated with any other type of weather related disasters.

TCs are associated with multiple hazards, including strong winds, storm surges, heavy rainfall and flooding. While the effects of winds and surge are mostly felt along the coastal areas near the landfall location, heavy rainfall and flooding are responsible for significant damage over much larger areas, even hundreds of kilometers from the coast. More than 50% of the fatalities associated with TCs between 1970 and 2004 were caused by fresh water flooding (http://www.nws.noaa.gov/os/water/ahps/pdfs/InlandFloodBrochure7F.pdf). Over the period 1963–2012, Rappaport (2014) showed that almost 50% of the U.S landfalling TCs have at least one fatality related to rain. Hurricane Ivan (2004) alone accounted for two-thirds of the total flood insurance payments made by the federal government in that year, impacting 23 different states (Czajkowski et al., 2013).

U.S. landfalling TCs are responsible for major flood events over large areas east of the Rocky Mountains, in particular along the eastern and central United States and along the coastal regions on the Gulf of Mexico (Villarini and Smith, 2010, Villarini and Smith, 2013, Villarini et al., 2011, Villarini et al., 2014). Although precipitation directly associated with TCs is less than 25% of the annual precipitation even in the most affected regions, the impacts can be extremely significant (e.g., Kunkel et al., 2010, Jiang and Zipser, 2010, Barlow, 2011).

Despite these negative socio-economic impacts, landfalling TCs have also been found to play a significant role as “drought busters” (e.g., Elsberry, 2002, Maxwell et al., 2012, Maxwell et al., 2013, Kam et al., 2013). Torrential rainfall associated with TCs occasionally can have the effect of breaking a prolonged drought by recharging reservoirs and elevating soil moisture. In these situations TC rainfall mitigates one environmental stressor, even as the potential for damage associated with the extreme rainfall, high-speed wind and ocean surge remain (e.g., Kam et al., 2013, Maxwell et al., 2013, Khouakhi and Villarini, 2016). Because rainfall associated with TCs has both significant positive and negative impacts on our society, it is critical that we understand how skillful current forecasting systems are in predicting rainfall associated with these storms to help us improve our preparedness and mitigation efforts.

Numerical Weather Prediction (NWP) models provide forecasts of a number of weather-related variables (e.g., precipitation, temperature at different levels) for different lead-times (e.g., Lorenc, 1986, Bougeault et al., 2010). However, quantitative information about the skill of NWP models in forecasting TC rainfall is still limited (Marchok et al., 2007, Mohanty et al., 2014). For a skillful prediction of TC rainfall, the models must predict the strength and distribution of the rainfall rate and wind fields together with the track and intensity of the TC system (see Halperin et al. (2013) for a discussion on the genesis forecasting of North Atlantic TCs). Therefore, precipitation forecasts from NWP models in general, and for TCs in particular, are inherently uncertain and subject to three types of error: localization, timing and intensity of precipitation events (e.g., Marchok et al., 2007). In this study, our goal is to evaluate the skill of NWP models in forecasting TC rainfall by quantifying their errors with respect to a reference (rain gauge-based) dataset. Moreover, five additional “observational” (remote sensing-based) datasets are also compared to the reference dataset: the skill of the NWP models in forecasting TC rainfall is quantified for different lead-times, and discussed and interpreted with respect to the performance of these “observational” products.

In this paper, the description of data and methodology is provided in Section 2, followed by results and discussion in Section 3. Section 4 summarizes the main points of the study and concludes the paper.