Heidke Skill Score Research Articles

Objective Classification of Tornadic and Nontornadic Severe Weather Outbreaks

Abstract Tornadoes often strike as isolated events, but many occur as part of a major outbreak of tornadoes. Nontornadic outbreaks of severe convective storms are more common across the United States but pose different threats than do those associated with a tornado outbreak. The main goal of this work is to distinguish between significant instances of these outbreak types objectively by using statistical modeling techniques on numerical weather prediction output initialized with synoptic-scale data. The synoptic-scale structure contains information that can be utilized to discriminate between the two types of severe weather outbreaks through statistical methods. The Weather Research and Forecast model (WRF) is initialized with synoptic-scale input data (the NCEP–NCAR reanalysis dataset) on a set of 50 significant tornado outbreaks and 50 nontornadic severe weather outbreaks. Output from the WRF at 18-km grid spacing is used in the objective classification. Individual severe weather parameters forecast by the model near the time of the outbreak are analyzed from simulations initialized at 24, 48, and 72 h prior to the outbreak. An initial candidate set of 15 variables expected to be related to severe storms is reduced to a set of 6 or 7, depending on lead time, that possess the greatest classification capability through permutation testing. These variables serve as inputs into two statistical methods, support vector machines and logistic regression, to classify outbreak type. Each technique is assessed based on bootstrap confidence limits of contingency statistics. An additional backward selection of the reduced variable set is conducted to determine which variable combination provides the optimal contingency statistics. Results for the contingency statistics regarding the verification of discrimination capability are best at 24 h; at 48 h, modest degradation is present. By 72 h, the contingency statistics decline by up to 15%. Overall, results are encouraging, with probability of detection values often exceeding 0.8 and Heidke skill scores in excess of 0.7 at 24-h lead time.

Preliminary Development and Evaluation of Lightning Jump Algorithms for the Real-Time Detection of Severe Weather

Abstract Previous studies have demonstrated that rapid increases in total lightning activity (intracloud + cloud-to-ground) are often observed tens of minutes in advance of the occurrence of severe weather at the ground. These rapid increases in lightning activity have been termed “lightning jumps.” Herein, the authors document a positive correlation between lightning jumps and the manifestation of severe weather in thunderstorms occurring across the Tennessee Valley and Washington D.C. A total of 107 thunderstorms from the Tennessee Valley; Washington, D.C.; Dallas, Texas; and Houston, Texas, were examined in this study. Of the 107 thunderstorms, 69 thunderstorms fall into the category of nonsevere and 38 into the category of severe. From the dataset of 69 isolated nonsevere thunderstorms, an average, peak, 1-min flash rate of 10 flashes per minute was determined. A variety of severe thunderstorm types were examined for this study, including a mesoscale convective system, mesoscale convective vortex, tornadic outer rainbands of tropical remnants, supercells, and pulse severe thunderstorms. Of the 107 thunderstorms, 85 thunderstorms (47 nonsevere, 38 severe) were from the Tennessee Valley and Washington, D.C., and these 85 thunderstorms tested six lightning jump algorithm configurations (Gatlin, Gatlin 45, 2σ, 3σ, Threshold 10, and Threshold 8). Performance metrics for each algorithm were then calculated, yielding encouraging results from the limited sample of 85 thunderstorms. The 2σ lightning jump algorithm had a high probability of detection (POD; 87%), a modest false-alarm rate (FAR; 33%), and a solid Heidke skill score (0.75). These statistics exceed current NWS warning statistics with this dataset; however, this algorithm needs further testing because there is a large difference in sample sizes. A second and more simplistic lightning jump algorithm named the Threshold 8 lightning jump algorithm also shows promise, with a POD of 81% and a FAR of 41%. Average lead times to severe weather occurrence for these two algorithms were 23 min. The overall goal of this study is to advance the development of an operationally applicable jump algorithm that can be used with either total lightning observations made from the ground, or in the near future from space using the Geostationary Operational Environmental Satellite Series R (GOES-R) Geostationary Lightning Mapper.

Severe local convective storms in Bangladesh: Part II.: Environmental conditions

This paper examines the environmental conditions of severe local convective storms during the pre-monsoon season (from March to May) in Bangladesh. We compared composite soundings on severe local convective storm days (SLCSD) with those on non-severe local convective storm days (NSLCSD) using rawinsonde data at 06 Bangladesh Standard Time (BST) in Dhaka (90.3°E and 23.7°N). Temperatures are rising in the lower layer and falling in the middle layer, and the amount of water vapor is significantly increasing in the lowest layer with southerly wind intensified on SLCSD compared with NSLCSD. This situation produces great thermal instability in the atmosphere on SLCSD. Convective parameters on SLCSD are computed with the rawinsonde data at 06 BST in Dhaka and compared with those on NSLCSD. The comparison shows that while most convective parameters related to thermal instability can discriminate between SLCSD and NSLCSD with statistical significance, no convective parameters related to the vertical wind shear can distinguish between the two categories. We evaluated the forecast skill of the convective parameters using Heidke Skill Score (HSS). The evaluation shows that the HSS for the Lifted Index and Precipitable Water are better among all parameters and have great forecast ability.

Multispectral remote sensing for rainfall detection and estimation at the source of the Blue Nile River

Most remote sensing based rainfall products have spatial resolutions ≥0.25° and temporal resolutions ≥1 day which are coarser than what is typically needed in hydrology. In this study, satellite data obtained from the precipitation radar (PR) of the Tropical Rainfall Measuring Mission (TRMM) which acquires data at 5km resolution once or twice a day and from the Spinning Enhanced Visible and Infrared Imager (SEVIRI) of the Meteosat Second Generation (MSG-2) which acquires data at 3km resolution at 15min interval. We evaluated three MSG-2 channels for rainfall detection in the Upper Blue Nile area in Ethiopia by the following indices: (1) 10.8μm brightness temperature, (2) rate of change of the 10.8μm brightness temperature, (3) space gradient of the 10.8μm brightness temperature, (4) brightness temperature difference (BTD) at the 10.8 and 6.2μm and (5) BTD at the 10.8 and 12.0μm channels. The evaluation was made through categorical statistics that are bias, probability of detection, false alarm ratio and Heidke skill score. In this work also, an exponential model was developed for thermal infrared based rainfall estimation. The model was evaluated using observations from a rain gauge network that we installed at the source of the Blue Nile River in Ethiopia.

Evaluation of heavy rainfall warning over India during summer monsoon season

India Meteorological Department (IMD) issues heavy rainfall warning for a meteorological sub-division when the expected 24 hours rainfall over any rain gauge station in that sub-division is likely to be 64.5 mm or more. Though these warnings have been provided since the inception of IMD, a few attempts have been made for quantitative evaluation of these warnings. Hence, a study is undertaken to verify the heavy rainfall warnings over 36 meteorological sub-divisions of India during monsoon months (June-September) and season as a whole. For this purpose, data of recent 5 years (2002-2006) has been taken into consideration. In this connection, the day when heavy rainfall is recorded over at least one station in a sub-division, has been considered as a heavy rainfall day for that sub-division.
 
 There is a large spatial and temporal variability in skill scores of heavy rainfall warnings over India during summer monsoon season. Considering the monsoon season as a whole, the Heidke Skill Score (HSS) is relatively less (<0.20) over the regions with less frequent heavy rainfall like Lakshadweep, southeast peninsula, Vidarbha, Marathwada, Jammu & Kashmir, Arunachal Pradesh and Nagaland, Manipur, Mizoram & Tripura (NMMT). It is higher (> 0.50) over Konkan & Goa, Madhya Maharashtra and Gujarat region. There has been improvement in the forecast skill with gradual increase in the critical success index and Heidke skill score over the years mainly due to the Numerical Weather Prediction (NWP) models' guidance available to the forecasters. However, the false alarm rate and missing rate are still very high (> 0.50), especially over many sub-divisions of northwest India, southeast peninsula and NMMT.

Support vector machines for spatiotemporal tornado prediction

The use of support vector machines (SVMs) for predicting the location and time of tornadoes is presented. In this paper, we extend the work by Lakshmanan et al. (Proceedings of 2005 IEEE international joint conference on neural networks (Montreal, Canada), 3, 2005a, 1642–1647) to use a set of 33 storm days and introduce some variations that improve the results. The goal is to estimate the probability of a tornado event at a particular spatial location within a given time window. We utilize a least-squares methodology to estimate shear, quality control of radar reflectivity, morphological image processing to estimate gradients, fuzzy logic to generate compact measures of tornado possibility and SVM classification to generate the final spatiotemporal probability field. On the independent test set, this method achieves a Heidke's skill score of 0.60 and a critical success index of 0.45.

Downscaling large-scale NCEP CFS to resolve fine-scale seasonal precipitation and extremes for the crop growing seasons over the southeastern United States

Seasonally predicted precipitation at a resolution of 2.5° was statistically downscaled to a fine spatial scale of ~20 km over the southeastern United States. The downscaling was conducted for spring and summer, when the fine-scale prediction of precipitation is typically very challenging in this region. We obtained the global model precipitation for downscaling from the National Center for Environmental Prediction/Climate Forecast System (NCEP/CFS) retrospective forecasts. Ten member integration data with time-lagged initial conditions centered on mid- or late February each year were used for downscaling, covering the period from 1987 to 2005. The primary techniques involved in downscaling are Cyclostationary Empirical Orthogonal Function (CSEOF) analysis, multiple regression, and stochastic time series generation. Trained with observations and CFS data, CSEOF and multiple regression facilitated the identification of the statistical relationship between coarse-scale and fine-scale climate variability, leading to improved prediction of climate at a fine resolution. Downscaled precipitation produced seasonal and annual patterns that closely resemble the fine resolution observations. Prediction of long-term variation within two decades was improved by the downscaling in terms of variance, root mean square error, and correlation. Relative to the coarsely resolved unskillful CFS forecasts, the proposed downscaling drove a significant reduction in wet biases, and correlation increased by 0.1–0.5. Categorical predictability of seasonal precipitation and extremes (frequency of heavy rainfall days), measured with the Heidke skill score (HSS), was also improved by the downscaling. For instance, domain averaged HSS for two category predictability by the downscaling are at least 0.20, while the scores by the CFS are near zero and never exceed 0.1. On the other hand, prediction of the frequency of subseasonal dry spells showed limited improvement over half of the Georgia and Alabama region.

Scale Interactions in Radar Rainfall Estimation Uncertainty

This study formulates an experimental strategy that couples rainfall observations from a locally deployed mobile X-band dual polarization radar with recording rain gauge measurements to create an error assessment of the spatio-temporal variability of high-resolution gridded rainfall fields. We discuss the trade-offs and values in the radar-rainfall uncertainty associated with different measurement resolutions. Overall, we found that radar measurement errors naturally decrease with averaging in space and time, but there is a generally negative value associated with the increase in spatial scale from 300 m to 5 km. Spatial averaging of radar rainfall increased the probability of detection slightly by 4%. However the decrease in relative root-mean-squared error (R-RMSE) was negligible as the false alarm rate increased by 10% and the Heidke skill score reduced by 16%. In terms of reducing measurement uncertainty, there was greater overall value by averaging temporally up to 60 min than averaging spatially to ...

Four Years of Tropical ERA-40 Vorticity Maxima Tracks. Part II: Differences between Developing and Nondeveloping Disturbances

Abstract Using a subset of the relative vorticity maxima (VM) tracks described in Part I, large-scale environmental fields, cold cloud area, and rainfall area are used to discriminate between developing and nondeveloping tropical disturbances in the eastern North Pacific (EPAC) and Atlantic Oceans. By using a minimum cold cloud coverage requirement, the nondeveloping VM are limited to disturbances with enhanced low-level relative vorticity and widespread deep convection. Linear discriminant analysis is used to determine the overall discrimination and the relative importance of each predictor for each basin separately. It is important to distinguish the two basins because, for many predictors, the differences between the basins are greater than the differences between developing and nondeveloping VM in each basin. Using the parametric forecast method, there is greater discrimination and prediction skill in the EPAC than in the Atlantic. There are also significant differences between the two basins in terms of the degree of discrimination provided by each of the predictors. Surprisingly, the mean vertical wind shear magnitude is greater for EPAC developing VM than for EPAC nondeveloping VM. Incorporating the satellite-derived predictors marginally improves the potential forecast skill in the EPAC but not in the Atlantic. The prediction skill (Heidke skill score) of tropical cyclogenesis in the Atlantic is similar to what has been obtained in previous studies using cloud cluster tracks. There is greater predictive skill in the EPAC.

Evaluation of a Satellite Multispectral VIS–IR Daytime Statistical Rain-Rate Classifier and Comparison with Passive Microwave Rainfall Estimates

Abstract A daytime surface rain-rate classifier, based on artificial neural networks (ANNs), is proposed for the Spinning Enhanced Visible and Infrared Imager (SEVIRI) on board the Meteosat-8 geostationary satellite. It is developed over the British Isles and surrounding waters, where the Met Office radar network provided the “ground precipitation truth” for training and validation. The algorithm classifies rain rate in five classes at 15 min and 5 km of time and spatial resolution, and is applied on daytime hours in a summer and winter database. A further ANN application is restricted to hours between 1200 and 1400 UTC for which the Advanced Microwave Scanning Radiometer for Earth Observing System (AMSR-E) on board the Aqua polar-orbiting satellite scans the U.K. area: ANN-classifier algorithms for the SEVIRI and AMSR-E data have been developed and the results have been compared. A reliable validation procedure is adopted to quantify the performance in view of the operational application of the daytime classifier and to investigate the relative skills of passive microwave and visible–infrared radiances in sensing precipitation if processed with equivalent algorithms. The key statistical parameters used are the equitable threat score (ETS) and the bias for rain–no rain classes and the Heidke skill score (HSS) for rain-rate classes. The SEVIRI daytime classifier shows, for mean seasonal conditions, the best performance in summer, with ETS = 47% and HSS = 22%, and in winter ETS = 36% and HSS = 17% were found. The comparison between AMSR-E and SEVIRI noon classifiers reveals a similar overall skill: in detecting rain areas, SEVIRI is slightly better than AMSR-E, while the opposite is true for rain-rate classification.

Precipitation process and rainfall intensity differentiation using Meteosat Second Generation Spinning Enhanced Visible and Infrared Imager data

A new day and night technique for precipitation process separation and rainfall intensity differentiation using the Meteosat Second Generation Spinning Enhanced Visible and Infrared Imager is proposed. It relies on the conceptual design that convective clouds with higher rainfall intensities are characterized by a larger vertical extension and a higher cloud top. For advective‐stratiform precipitation areas, it is assumed that areas with a higher cloud water path (CWP) and more ice particles in the upper parts are characterized by higher rainfall intensities. First, the rain area is separated into areas of convective and advective‐stratiform precipitation processes. Next, both areas are divided into subareas of differing rainfall intensities. The classification of the convective area relies on information about the cloud top height gained from water vapor‐IR differences and the IR cloud top temperature. The subdivision of the advective‐stratiform area is based on information about the CWP and the particle phase in the upper parts. Suitable combinations of temperature differences (ΔT3.9–10.8, ΔT3.9–7.3, ΔT8.7–10.8, ΔT10.8–12.1) are incorporated to infer information about the CWP during nighttime, while a visible and a near‐IR channel are considered during the daytime. ΔT8.7–10.8 and ΔT10.8–12.1 are particularly included to supply information about the cloud phase. Intensity differentiation is realized by using pixel‐based confidences for each subarea calculated as a function of the respective value combinations of the previously mentioned variables. For the calculation of the confidences, the value combinations are compared with ground‐based radar data. The proposed technique is validated against ground‐based radar data and shows an encouraging performance (Heidke skill score 0.07–0.2 for 15‐min intervals).

A Statistical Evaluation of GOES Cloud-Top Properties for Nowcasting Convective Initiation

Abstract The goal of this project is to validate and extend a study by Mecikalski and Bedka that capitalized on information the Geostationary Operational Environmental Satellite (GOES) instruments provide for nowcasting (i.e., 0–1-h forecasting) convective initiation through the real-time monitoring of cloud-top properties for moving cumuli. Convective initiation (CI) is defined as the first occurrence of a ≥35-dBZ radar echo from a cumuliform cloud. Mecikalski and Bedka’s study concluded that eight infrared GOES-based “interest fields” of growing cumulus clouds should be monitored over 15–30-min intervals toward predicting CI: the transition of cloud-top brightness temperature to below 0°C, cloud-top cooling rates, and instantaneous and time trends of channel differences 6.5–10.7 and 13.3–10.7 μm. The study results are as follows: 1) measures of accuracy and uncertainty of Mecikalski and Bedka’s algorithm via commonly used skill scoring procedures, and 2) a report on the relative importance of each interest field to nowcasting CI using GOES. It is found that for nonpropagating convective events, the skill scores are dependent on which CI interest fields are considered per pixel and are optimized when three–four fields are met for a given 1-km GOES pixel in terms of probability of detection, and threat and Heidke skill scores. The lowest false-alarm rates are found when one field is used: that associated with cloud-top glaciation 30 min prior to CI. Subsequent recommendations for future research toward improving Mecikalski and Bedka’s study are suggested especially with regard to constraining CI nowcasts when inhibiting factors are present (e.g., capping inversions).

Developments in Operational Long-Range Climate Prediction at CPC

Abstract The science, production methods, and format of long-range forecasts (LRFs) at the Climate Prediction Center (CPC), a part of the National Weather Service’s (NWS’s) National Centers for Environmental Prediction (NCEP), have evolved greatly since the inception of 1-month mean forecasts in 1946 and 3-month mean forecasts in 1982. Early forecasts used a subjective blending of persistence and linear regression-based forecast tools, and a categorical map format. The current forecast system uses an increasingly objective technique to combine a variety of statistical and dynamical models, which incorporate the impacts of El Niño–Southern Oscillation (ENSO) and other sources of interannual variability, and trend. CPC’s operational LRFs are produced each midmonth with a “lead” (i.e., amount of time between the release of a forecast and the start of the valid period) of ½ month for the 1-month outlook, and with leads ranging from ½ month through 12½ months for the 3-month outlook. The 1-month outlook is also updated at the end of each month with a lead of zero. Graphical renderings of the forecasts made available to users range from a simple display of the probability of the most likely tercile to a detailed portrayal of the entire probability distribution. Efforts are under way at CPC to objectively weight, bias correct, and combine the information from many different LRF prediction tools into a single tool, called the consolidation (CON). CON ½-month lead 3-month temperature (precipitation) hindcasts over 1995–2005 were 18% (195%) better, as measured by the Heidke skill score for nonequal chances forecasts, than real-time official (OFF) forecasts during that period. CON was implemented into LRF operations in 2006, and promises to transfer these improvements to the official LRF. Improvements in the science and production methods of LRFs are increasingly being driven by users, who are finding an increasing number of applications, and demanding improved access to forecast information. From the forecast-producer side, hope for improvement in this area lies in greater dialogue with users, and development of products emphasizing user access, input, and feedback, including direct access to 5 km × 5 km gridded outlook data through NWS’s new National Digital Forecast Database (NDFD).

Terminal aerodrome forecast verification in Austro Control using time windows and ranges of forecast conditions

AbstractTerminal aerodrome forecasts (TAFs) are widely used meteorological forecasts for flight planning. Therefore, there is considerable interest in assessing their accuracy, skill and value. TAFs give information about the expected conditions of wind, visibility, significant weather and clouds at airports. Using different types of change groups, the forecaster gives a range of possible values valid for a time interval, the shortest interval being 1 h. A TAF thus contains a range of forecast conditions for each hour.Point verification has proved to be difficult for TAFs. To ease these difficulties, time and meteorological state constraints are relaxed in the method described in this paper. This is done by verifying two conditions for each hour of the TAF. The highest (or most favourable) observed value is used to score the highest forecast value, and the lowest (or most adverse) observed value is used to score the lowest forecast value. Entries are made accordingly into two contingency tables. The contingency tables are specific for weather element and lead time.Verification results should give feedback to forecasters. Contingency tables show the strengths and weaknesses of TAF, and displays for individual TAFs are available in the sense of ‘eyeball verification’. For management information, common verification measures for categorical events (such as the Gerrity Score and the Heidke Skill Score) are calculated from the contingency tables. For answering specific customer questions, verification results in respect to certain values of weather elements are available. Copyright © 2008 Royal Meteorological Society

A Remote Sensed Data Combined Method for Sea Fog Detection

Steam and advection fogs are frequently observed in the Yellow Sea from March to July except for May. This study uses remote sensing (RS) data for the monitoring of sea fog. Meteorological data obtained from the Ieodo Ocean Research Station provided a valuable information for the occurrence of steam and advection fogs as a ground truth. The RS data used in this study were GOES-9, MTSAT-1R images and QuikSCAT wind data. A dual channel difference (DCD) approach using IR and shortwave IR channel of GOES-9 and MTSAT-1R satellites was applied to detect sea fog. The results showed that DCD, texture-related measurement and the weak wind condition are required to separate the sea fog from the low cloud. The QuikSCAT wind data was used to provide the wind speed criteria for a fog event. The laplacian computation was designed for a measurement of the homogeneity. A new combined method, which includes DCD, QuikSCAT wind speed and laplacian computation, was applied to the twelve cases with GOES-9 and MTSAT-1R. The threshold values for DCD, QuikSCAT wind speed and laplacian are -2.0 K, 8 m s-1 and 0.1, respectively. The validation results showed that the new combined method slightly improves the detection of sea fog compared to DCD method: improvements of the new combined method are 5 ~ 6 % increases in the Heidke skill score, 10% decreases in the probability of false detection, and 30 ~ 40% increases in the odd ratio.

A Fuzzy Logic–Based Analog Forecasting System for Ceiling and Visibility

Abstract WIND-3 is an application for aviation weather forecasting that uses the analog method to produce deterministic predictions of cloud ceiling height and horizontal visibility at airports. For data, it uses historical and current airport observations [routine aviation weather reports (METARs)], and model-based guidance. It uses the perfect prognosis assumption as it is designed to use any model-based predictions of wind direction and speed, temperature and humidity, and precipitation occurrence and type to specify conditions in the 1–24-h projection period. To identify and rank analogs, according to their degree of similarity with the present situation, it uses a fuzzy logic–based algorithm to measure similarity between past situations, which are complete series of METARs, and current situations, which are a composite of recent METARs and model-based guidance. It uses the retrieved analog ensemble, the set of most similar analogs, to make predictions of ceiling and visibility in the 1–24-h projection period. WIND-3 has been tested by being run continuously in real time for 1 yr, producing forecasts for 190 major Canadian airports. It produces accurate forecasts, based on summaries of Heidke skill score (HSS) statistics, and compared to two benchmarks, persistence and official aerodrome forecasts [terminal aerodrome forecasts (TAFs)]. WIND-3 predictions of instrument flight regulation (IFR) conditions in the 0–6-h period have an HSS of 0.56, and in the 7–24-h period have an HSS of about 0.40, compared to respective HSS scores for persistence forecasts of 0.53 and less than 0.20.

Short-term forecast of thunderstorms in Argentina

The eastern Andes have been described as a hot spot in severe convection. The damages caused in that particular area by convection are much higher than in other latitudes. In contrast, very few studies have been conducted as yet on the thermodynamic characterization of these events or on their forecasting. This paper assesses pre-convective conditions in south-western Argentina on the basis of radiosonde data of 713 days. Additionally, an objective short-term forecast model for storms has been set up. The linear discriminant analysis provided information about the likelihood of storms in the study zone. The forward stepwise method was used to set up the discriminant equation. Testing the results on a validation sample, the outcome was: Probability of detection (POD) = 0.814, False Alarm Rate (FAR) = 0.200; Heidke Skill Score (HSS) = 0.535. Finally, a forecast model was developed for severe storms too, but the results achieved were not so satisfactory.

The skill of convective parameters and indices to predict isolated and severe thunderstorms

Abstract. The preconvective environment on days with ordinary, widespread, and severe thunderstorms in Southwest Germany was investigated. Various thermodynamic and kinetic parameters calculated from radiosoundings at 12:00 UTC were verified against subsequent thunderstorm observations derived from SYNOP station data, radar data, and damage reports of a building insurance company. The skill of the convective parameters and indices to predict thunderstorms was evaluated by means of probability distribution functions, probabilities of thunderstorms according to an index threshold, and skill scores like the Heidke Skill Score (HSS) that are based on categorical verification. For the ordinary decision as to whether a thunderstorm day was expected or not, the best results were obtained with the original Lifted Index (80% prediction probability for LI≤−1.73; HSS=0.57 for LI≤1.76), the Showalter Index, and the modified K-Index. Considering days with isolated compared to widespread thunderstorms, the best performance is reached by the Deep Convective Index. For days with severe thunderstorms that caused damage due to hail, local storms or floods, the best prediction skill is found again for the Lifted Index and the Deep Convective Index, but also for the Potential Instability Index, the Delta-θe Index, and a version of the CAPE, where the lifting profile is determined by averaging over the lowest 100 hPa.

다중 선형 회귀를 이용한 PNU/CME CGCM의 동아시아 여름철 강수예측 보정 연구

강수는 다양한 대기 변수들의 영향으로 나타나기 때문에 비선형성이 매우 강하다. 따라서 역학 모형을 통해 예측된 강수의 보정은 비선형 모형인 인공 신경망 등을 통해 가능할 것이지만, 인공 신경망의 경우 초기 가중치 선택, 지역 최소화 문제, 뉴런의 수 결정 등의 문제로 인한 한계가 있다. 그러므로 본 연구에서는 가장 보편적으로 사용되는 다중 선형 회귀 모형을 이용하여 CGCM에 의해 모사된 강수를 보정하였으며, 예측성을 살펴보았다. 이를 위하여 우선 PNU/CME 접합 대순환 모형(Coupled General Circulation model, CGCM)(박혜선과 안중배, 2004)을 이용하여 1979년부터 2005년까지 매해 4월부터 8월까지 5개월간 앙상블 적분을 하였다. 적분 결과 중 한반도를 포함한 동북아시아 지역<TEX>$(110^{\circ}E-145^{\circ}E,\;25^{\circ}N-55^{\circ}N)$</TEX>의 여름철인 6월(리드 2), 7월(리드 3), 8월(리드 4) 및 여름철 평균인 JJA(from June to August) 기간의 PNU/CME CGCM에 의해 모사된 강수를 보정하기 위해 다중 선형 회귀(Multiple Linear Regression, MLR)를 이용하였다. PNU/CME 접합 대순환 모형의 결과 중 강수, 500 hPa 연직 속도, 200 hPa 발산장, 지상 기온 등의 예측 인자와 관측 강수와의 선형적인 관계를 이용하여 MLR 모형을 구축하였다. 그리고 교차 검증(cross- validation)을 수행하여 PNU/CME 접합 대순환 모형의 결과와 교차 검증 결과를 비교하였다. 상관계수, 적중률 (hit rate), 오보율(false alarm rate) 그리고 Heidke 기술 점수(Heidke skill score) 등을 살펴본 바, 보정하지 않은 모형의 결과에 비해 MLR 모형을 이용하여 보정한 결과의 강수에 대한 예측성이 뛰어난 것을 알 수 있었다. Because precipitation is influenced by various atmospheric variables, it is highly nonlinear. Although precipitation predicted by a dynamic model can be corrected by using a nonlinear Artificial Neural Network, this approach has limits such as choices of the initial weight, local minima and the number of neurons, etc. In the present paper, we correct simulated precipitation by using a multiple linear regression (MLR) method, which is simple and widely used. First of all, Ensemble hindcast is conducted by the PNU/CME Coupled General Circulation Model (CGCM) (Park and Ahn, 2004) for the period from April to August in 1979-2005. MLR is applied to precipitation simulated by PNU/CME CGCM for the months of June (lead 2), July (lead 3), August (lead 4) and seasonal mean JJA (from June to August) of the Northeast Asian region including the Korean Peninsula <TEX>$(110^{\circ}-145^{\circ}E,\;25-55^{\circ}N)$</TEX>. We build the MLR model using a linear relationship between observed precipitation and the hindcasted results from the PNU/CME CGCM. The predictor variables selected from CGCM are precipitation, 500 hPa vertical velocity, 200 hPa divergence, surface air temperature and others. After performing a leave-oneout cross validation, the results are compared with the PNU/CME CGCM's. The results including Heidke skill scores demonstrate that the MLR corrected results have better forecasts than the direct CGCM result for rainfall.

Probabilistic forecast of daily areal precipitation focusing on extreme events

Abstract. A dynamical downscaling scheme is usually used to provide a short range flood forecasting system with high-resolved precipitation fields. Unfortunately, a single forecast of this scheme has a high uncertainty concerning intensity and location especially during extreme events. Alternatively, statistical downscaling techniques like the analogue method can be used which can supply a probabilistic forecasts. However, the performance of the analogue method is affected by the similarity criterion, which is used to identify similar weather situations. To investigate this issue in this work, three different similarity measures are tested: the euclidean distance (1), the Pearson correlation (2) and a combination of both measures (3). The predictor variables are geopotential height at 1000 and 700 hPa-level and specific humidity fluxes at 700 hPa-level derived from the NCEP/NCAR-reanalysis project. The study is performed for three mesoscale catchments located in the Rhine basin in Germany. It is validated by a jackknife method for a period of 44 years (1958–2001). The ranked probability skill score, the Brier Skill score, the Heidke skill score and the confidence interval of the Cramer association coefficient are calculated to evaluate the system for extreme events. The results show that the combined similarity measure yields the best results in predicting extreme events. However, the confidence interval of the Cramer coefficient indicates that this improvement is only significant compared to the Pearson correlation but not for the euclidean distance. Furthermore, the performance of the presented forecasting system is very low during the summer and new predictors have to be tested to overcome this problem.