Urban extreme rainfall events: categorical skill of WRF model simulations for localized and non‐localized events

Abstract

An objective method is used for determining the rainfall threshold for identifying extreme rainfall events (EREs) over the urban city, Bangalore, using observed rainfall data for a period of 35 years (1971–2005). Using this threshold, 52 EREs were identified during the period 2010–2014 using high‐resolution rain‐gauge observations. From these EREs, 15 localized and non‐localized events were chosen based on spatial distribution to examine the forecast skill of the Weather Research and Forecasting (WRF) model. Apart from the conventional verification methods, a number of skill scores and indices were defined for a comprehensive evaluation of rainfall model skill. In general, the forecast underpredicted the magnitude of localized and non‐localized EREs for the majority of cases; however, the model overpredicted light rainfall (≤10 mm day−1). The model showed a success rate of 59% in simulating light rainfall for localized EREs while 12% of events were missed and 29% were wrongly predicted. The success rate was significantly reduced at higher rainfall categories for localized and non‐localized EREs, where the forecast missed the majority of rainfall events. The Reliability Index (RI) computed clearly showed that model skill is relatively higher for non‐localized EREs compared to localized EREs. The average forecast reliability for non‐localized and localized EREs were 74 and 51%, respectively. For localized EREs, model skill is relatively higher in rainfall location prediction (61%) compared to area (44%) and intensity (46%) prediction; while in the case of non‐localized EREs, model skill is similar for location, intensity and area prediction. It is found that coupling an urban canopy model with WRF reduces the model errors particularly for lower rainfall thresholds.