傳統的雙偏極化雷達定量降水估計(Radar Quantitative Precipitation Estimation， RQPE)通常依賴基於區域統計的雨滴粒徑分布(Raindrop Size Distribution， RSD)所定義的多項式降雨估計式。然而因為雨滴粒徑分布(RSD)在時間與空間上的變化，使得這些多項式為高度非線性。此外，在複雜地形上，雷達定量降雨(RQPE)會受到部分波束遮擋以及因波束本身與地表的高度差，出現雲微物理變化造成估計降雨的誤差，因此對雷達定量降雨(RQPE)在垂直方向上進行修正對於真實估計地表降水來說至關重要。其中為解決雷達定量降水估計(RQPE)固有的複雜性與非線性問題，本研究使用深度學習(Deep Learning， DL)方法對雙偏極化雷達參數測量值(DPMs， i.e.，Z_HH，Z_DR，K_DP)與降雨率(R)之間的關係建立模型，以取代傳統的冪律多項式。另外為了進一步解決地形帶來的挑戰，本研究開發垂直修正模型(TopoCorr-Net)結合多層的雙偏參數空間分布透過卷積神經擷取降水的垂直變化趨勢，得到降雨估計在高低層之間的變量。從梅雨個案的應用結果分析顯示，基於雷達觀測資料訓練的垂直修正模型(TopoCorr-Net)能夠有效捕捉估計降雨的垂直變化，並提升地表附近降雨率的估計精度。;Traditional radar quantitative precipitation estimation (RQPE) based on dual-polarization radar usually relies on power-law rainfall relations derived from regional raindrop size distribution (RSD) statistics. However， because RSD varies with time and location， these relations are highly nonlinear. In complex terrain， RQPE is further affected by partial beam blockage and vertical differences between radar observations and surface rainfall， leading to additional estimation errors. Therefore， correcting the vertical structure of rainfall is important for improving surface rainfall estimation. To address these issues， this study applies a deep learning (DL) approach to model the relationships among dual-polarization radar measurements， including reflectivity (ZH)， differential reflectivity (ZDR)， specific differential phase (KDP)， and rainfall rate (R)， thereby replacing traditional power-law equations. In addition， a vertical correction model， TopoCorr-Net， is developed to reduce terrain-related errors. The model combines multi-level radar parameter fields and utilizes convolutional neural networks to learn the vertical variation of precipitation and estimate rainfall differences between observation heights. Results from Mei-Yu rainfall events show that TopoCorr-Net can effectively capture vertical variations in rainfall and improve rainfall estimates near the surface. These results demonstrate the potential of combining deep learning and vertical correction techniques to enhance RQPE performance in complex terrain regions.
