Hazard at weather scale for extreme rainfall forecast reduces uncertainty

Abstract

Globally increasing intensity and frequency of extreme rainfall events demand reliable early warning systems. Despite significant improvements in the skills of weather models, the state-of-art extreme rainfall forecasts, at a sufficient lead time, still suffer from high biases, high uncertainties, low hit rates, and high false alarms. Bias correction methods often improve the performances of the models, but still, the skills remain moderate. Here, we propose a new methodology to forecast extreme rainfall events, in terms of hazard, instead of rainfall amount. At a weather scale, we define ‘hazard’ as the probability of occurrence of an extreme rainfall event, given a forecasted rainfall for a day with sufficient lead time. The conditional probability is obtained from the past observed data and the hindcast. The method is applied to India with observations from the India Meteorological Department (IMD) and hindcasts from the Global Ensemble Forecast System (GEFS) Reforecast Version 2 for 1985–2015. Extreme days at a grid level are defined as the days with observed rainfall exceeding the 95th percentile. Accordingly, we calculate the hazard for all the lead days till 15 days. For most of the extremes in each grid, the model can predict an extreme showing a high hazard value greater than 0.6 from lead day 7. This high hit rate may give the stakeholders adequate time to plan mitigation strategies. Comparing the proposed method with traditional methods, we find a significant improvement in terms of hit rate and the uncertainty across the ensembles.