The calibration method and the optimal probability threshold were proposed in order to improve the skill and utility of an ensemble-based Probability of Precipitation (PoP) forecast. The raw ensemble PoP forecasts were derived from the Japan Meteorolo...
The calibration method and the optimal probability threshold were proposed in order to improve the skill and utility of an ensemble-based Probability of Precipitation (PoP) forecast. The raw ensemble PoP forecasts were derived from the Japan Meteorological Agency (JMA) medium-range Ensemble Prediction System (EPS). In terms of 0.5, 1, 2, 3, 6, 12, 24, 48, and 72 mm day-1 precipitation thresholds, the 24-hour accumulated PoP forecasts were verified as a function of forecast time up to 9 days from January 2003 to December 2004.
Comparison between the raw and the calibrated PoP forecasts was carried out by using several skill measurements: reliability, the relative operating characteristic, and the Brier skill score. It is found that the raw PoP forecasts were evidently biased, exhibiting over-forecasting to the light rainfall events and under-forecasting to the heavy rainfall events. The 5-day calibrated PoP forecasts of the 1 mm day-1 precipitation threshold provide the excellent reliability close to the perfect reliability. In addition, the relative operating characteristics and the Brier skill scores also show that the calibrated ensemble forecasts are superior to the raw ensemble forecasts. In terms of the threshold of 48 mm day-1, the calibration provides less skill because of the rarity of the events screened by the threshold of 48 mm day-1.
Skill comparisons among the control forecast, the ensemble mean forecast, and the deterministic ensemble forecast selected by the optimal probability thresholds (perfect bias (bias=1) or maximum equitable threat score) were carried out. Both deterministic ensemble forecasts selected by optimal probability thresholds show better skill than the control and the ensemble mean forecasts. The monthly variations of the bias and the equitable threat scores for 5-day deterministic ensemble forecasts selected by two optimal probability thresholds are investigated from January 2003 to December 2004. It is found that the under-forecasting tendency is prevail to all precipitation thresholds and the tendency is stronger according to the increase of the precipitation threshold. The amplitude of the equitable threat score is decreasing as increasing the precipitation thresholds. In case of the thresholds of 1 and 6 mm day-1, the seasonal variability, that the equitable threat score decreases in summer and increases in winter, is evidently found. In case of the threshold of 48 mm day-1, the different seasonal variability, that the score becomes higher from March to June, is observed.
To take full advantage of the rare event data and to give the warning of them in advance, the potential index was developed on the basis of the climatology in terms of the observed precipitation and the calibrated PoP forecasts. The skill of the normalized potential index (NPI) was measured by the hit rate, the false alarm rate, and the forecast limit as a function of forecast days up to 9 days in January, April, July, and October 2003 and 2004. It is found that the normalized potential index can effectively describe the potential risk of the forecast event up to 9 days. It is also found that the hit rate is larger than the false alarm rate for 2 years. For more detail assessment, the frequency distribution of daily accumulated precipitation events is investigated. The relatively small frequency among the bins from 6 to 72 mm day-1 precipitation thresholds is found compared with other frequency distributions. In addition, the frequency distributions of the observed precipitation events in October 2003 and July 2004 also show the similar histograms.