本研究主要目標為發展雷達回波影像辨識技術，用以比對眾多系集成員的預報結果，從中客觀篩選 出模式預報與觀測回波接近的成員，並探討及評估其極短期（3 至12 小時內）定量降水預報的表現。研 究以兩種圖形辨識理論為基礎進行比對，分別為動差不變量理論和皮爾森相關係數，前者可將二維回波 圖形轉化為具物理幾何意義的七個分量，尤其特別發展片段逐步分塊的重複計算方式以提高辨識能力， 並利用相似度演算法來量化模式和觀測回波的相似程度，後者為統計上應用至平面空間的相關係數計算。 研究個案選取2013 年康芮颱風之8 月28 日1200 UTC 至29 日0600 UTC 期間，每三小時作為觀測 回波辨識時間，使用模式資料包括中央氣象局發展之Weather Research and Forecasting (WRF)區域模式 (WRFD、TWRF)和系集模式(WRF Ensemble Prediction System, WEPS)，在特定的比對策略設計下，系集 預報成員共有440 個。研究結果顯示模式雨量的公正預兆得分(Equitable Threat Score, ETS)隨著辨識排 名順序大致呈現線性遞減的趨勢，雖然名次間的ETS 仍有高低跳動的情況。以七個個案平均而言，辨識 排序前10 名的系集平均雨量之偏倚得分(Bias Score, BS)在中小雨量略呈現過度預報（BS 約在1.1 至 1.4），而較大雨量門檻則為預報不足之結果。辨識排序前10 名系集平均雨量之3 小時累積雨量在20 mm (50 mm)雨量門檻的ETS 約為0.47 (0.27)，而WRFD、TWRF 和WEPS 系集平均則分別為0.35 (0.32)、 0.17 (0.11)和0.19 (0.06)。除WRFD 在較大雨量門檻下表現較好外，辨識比對方法皆呈現明顯提升雨量 預報技術得分的結果。此外，12 小時累積雨量在130 mm 雨量門檻的ETS 亦由WEPS 系集平均的0.1， 經辨識篩選後，提升至約0.25。因此，透過此雷達回波影像辨識技術可有效挑選出與實際觀測相近的系 集預報成員，以期能進一步改善極短期定量降水預報之能力，並在預報作業上提供更即時且更有用的客 觀參考資訊。

The ensemble quantitative precipitation nowcasting (also referred to as short-range quantitative precipitation forecast) is developed and explored by objectively selecting the analogs to the latest observed radar reflectivity among numerous ensemble members based on two pattern recognition theories. One is called moment invariants theory that is able to extract seven components with specific geometric features, and the other is based on Pearson's correlation coefficient applied to compare two planar images. In particular, the repeated calculation of moment invariants within piecewise sections of an entire image is proposed to enhance the recognition capability. In addition, the similarity algorithm is adopted to quantify the degree of how one ensemble member resembles the observed radar reflectivity. Under a designed strategy, 440 members, which consist of two deterministic models (WRFD and TWRF) and the ensemble prediction system (WEPS) based on the Weather Research and Forecasting (WRF) regional model, are generated and compared to the radar observation for the case of Typhoon Kong-Rey in 2013. The results show that the equitable threat score (ETS) of accumulated rainfalls appears to have the linear decline trend following the similarity order. The ETS of 3-h accumulated rainfalls based on the ensemble mean of the first 10 ranks from two analog methods is about 0.47 (0.27) at the rainfall threshold of 20 mm (50 mm), which is higher than that based on the original WEPS ensemble mean with a ETS value of 0.19 (0.06). This indicates that the approach to finding similar radar images has better forecasting skill than the unrecognized ensemble output in terms of the short-term precipitation. Overall, the analog-based technique that objectively searches for forecast resemblances to the latest observation is shown to have the ability to improve the short-range quantitative precipitation forecast, thus adding value to the ensemble forecasts and providing the real-time guidance for the operational forecast.