Statistical Prediction of Heavy Rain in South Korea

This study is aimed at the development of a statistical model for forecasting heavy rain in South Korea. For the 3-hour weather forecast system, the 10 km×10 km area-mean amount of rainfall at 6 stations （Seoul, Daejeon, Gangreung, （Jwangju, Busan, and Jeju） in South Korea are used. And the corresponding 45 synoptic factors generated by the numerical model are used as potential predictors. Four statistical forecast models （linear regression model, logistic regression model, neural network model and decision tree model） for the occurrence of heavy rain are based on the model output statistics （MOS） method. They are separately estimated by the same training data. The thresholds are considered to forecast the occurrence of heavy rain because the distribution of estimated values that are generated by each model is too skewed. The results of four models are compared via Heidke skill scores. As a result, the logistic regression model is recommended.

Statistical Guidance on Seasonal Forecast of Korean Dust Days over South Korea in the Springtime

This study aimed to develop the seasonal forecast models of Korean dust days over South Korea in the springtime. Forecast mode was a ternary forecast （below normal, normal, above normal） which was classified based on the mean and the standard deviation of Korean dust days for a period of 30 years （1981-2010）. In this study, we used three kinds of monthly data： the Korean dust days observed in South Korea, the National Center for Environmental Prediction in National Center for Atmospheric Research （NCEP/NCAR） reanalysis data for meteorological factors over source regions of Asian dust, and the large-scale climate indices offered from the Climate Diagnostic Center and Climate Prediction Center in NOAA. Forecast guidance consisted of two components; ordinal logistic regression model to generate trinomial distributions, and conversion algorithm to generate ternary forecast by two thresholds. Forecast guidance was proposed for each month separately and its predictability was evaluated based on skill scores.

Ternary Forecast of Heavy Snowfall in the Honam Area, Korea

The objective of this study is to improve the statistical modeling for the ternary forecast of heavy snowfall in the Honam area in Korea. The ternary forecast of heavy snowfall consists of one of three values, 0 for less than 50 mm, 1 for an advisory （50-150 ram）, and 2 for a warning （more than 150 mm）. For our study, the observed daily snow amounts and the numerical model outputs for 45 synoptic factors at 17 stations in the Honam area during 5 years （2001 to 2005） are used as observations and potential predictors respectively. For statistical modeling and validation, the data set is divided into training data and validation data by cluster analysis. A multi-grade logistic regression model and neural networks are separately applied to generate the probabilities of three categories based on the model output statistic （MOS） method. Two models are estimated by the training data and tested by the validation data. Based on the estimated probabilities, three thresholds are chosen to generate ternary forecasts. The results are summarized in 3 × 3 contingency tables and the results of the three-grade logistic regression model are compared to those of the neural networks model. According to the model training and model validation results, the estimated three-grade logistic regression model is recommended as a ternary forecast model for heavy snowfall in the Honam area.

Guidance on the Choice of Threshold for Binary Forecast Modeling

This paper proposes useful guidance on the choice of threshold for binary forecasts. In weather forecast systems, the probabilistic forecast cannot be used directly when estimated too smoothly. In this case, the binary forecast, whether a meteorological event will occur or not, is preferable to the probabilistic forecast. A threshold is needed to generate a binary forecast, and the guidance in this paper encompasses the use of skill scores for the choice of threshold according to the forecast pattern. The forecast pattern consists of distribution modes of estimated probabilities, occurrence rates of observations, and variation modes. This study is performed via Monte-Carlo simulation, with 48 forecast patterns considered. Estimated probabilities are generated by random variate sampling from five distributions separately. Varying the threshold from 0 to 1, binary forecasts are generated by threshold. For the assessment of binary forecast models, a 2×2 contingency table is used and four skill scores （Heidke skill score, hit rate, true skill statistic, and threat score） are compared for each forecast pattern. As a result, guidance on the choice of skill score to find the optimal threshold is proposed.