降雨資料於水文模式中是一個關鍵因子，不同的降雨資料將影響模式的表現。本研究之研究區域為新竹鳳山溪流域面積約250平方公里。透過HEC-HMS模擬山區地表逕流，其中HEC採用SCS-CN的方式評估降雨損失，並匯入WASH123D以評估模擬一維河道與二維地表水文變化。此模式透過六場降雨事件驗證與校準後，顯示檢定結果校準之相關係數在0.8，均方根誤差為0.21至0.43m，且趨勢與觀測值相符合，判斷模式具有一定的能力用以模擬河道水位。 本研究為評估雷達降雨資料之正確性，故將此資料與雨量站降雨資料輸入於水文模式中模擬河川水位變化。結果表明雷達降雨推估(QPE)比雨量站資料更能反應出整個流域的變化。另一方面，雷達降雨預測(QPF)，在預測河川水位變化時，雖然能反應出水位的變化，但預測結果並不理想。故本研究透過物理性修正(RT)法與統計性修正(SVM)法來修正QPF預測結果，其結果表明，兩種修正方式皆可提升河川水位預測的準確性，其中RT之修正結果較佳，能精確地預測未來洪水的發生與轉折。而SVM之修正結果，雖能使整體水位變化更貼近觀測值，但仍有誤差存在，其原因可能為訓練資料不足或是特徵向量代表性不佳所致。 ;Rainfall is a key input data for flood forecasting but inherent errors in radar rainfall data can lead to poor performance of predictive hydrological models. A hydrometerological study in a Fengshan creek basin was conducted to check the feasibility of weather radar rainfall and flood forecasting. The study region has drainage area of 250 km2 located in Hsinchu, Taiwan. Rainfall-induced surface runoff from six rainstorm events were obtained from HEC-HMS using SCS-CN method. A coupled 1-D river and 2-D overland model, WASH123D was applied to simulate river stage. The performance of the model was satisfactory with correlation coefficient of 0.8 and root mean square error of 0.21 to 0.43m during calibration and validation, respectively. It was found that the simulated trends were consistent with the observed values, and WASH123D model has ability to simulate the water level of river. In order to evaluate the feasibility of rainfall radar data as input, the hydrological model was also simulated for river water level. It was found that model performs better when it uses radar rainfall estimation (QPE) as input rather than rainfall station data. Results of using one-hour Radar rainfall prediction (QPF) as input for water level forecasting were encouraging but have over and under estimation for different rainstorm events. Therefore, QPF prediction results were corrected through physical correction (RT) and statistical correction (SVM) methods. The results showed that both correction methods can improve the accuracy of river water level prediction. However, RT correction results are better and can accurately predict the occurrence and transition of future floods. while, the SVM correction results unable to predict the occurrence and transition of future floods accurately may be due to insufficient training data or poor representation of feature vectors.
