過去文獻已明確指出沿岸洪水溢淹的範圍可受多種環境因素影響,其中包含低窪地的地形條件、下游潮位、降雨型態以及上游流域的匯入流量等。本研究透過遞歸神經網路(Recurrent Neural Network, RNN)模型,配合水文和地貌因子的詳盡分析,以研發建置一套適用於沿海地區的有效淹水模擬方案。本研究的新穎之處乃在於利用每個網格的地形濕度指數(topographic wetness index, TWI),針對所有輸入資料進行初步分類以及各別訓練,以提高洪水模擬的整體準確性。研究中同時應用水動力方程式為基礎的數值淹水模式產製不同水文條件情況下之淹水圖資,以作為RNN模型練過程所需之目標淹水深度,進而透過機器學習技術分析下游地形、潮位、降雨強度與沿海溢淹於時間及空間分佈的相關性。本研究著重於評估此替代方案之實務可行性,藉由考量熱帶氣旋逼近所導致之風暴潮和極端降雨事件,以提昇預測沿海溢淹情資的效率和穩定性。本研究所開發的淹水模擬方式可作為數值淹水模型之替代方案,以強化颱洪期間執行預警模式的時效性與可靠性。
Coastal floodplain inundation is affected by a variety of environmental factors such as the rainfall pattern, tide elevation, topography of the low-lying land, inflow discharge from the upstream drainage area, etc. In this study, a creative methodology is developed for predicting the inundation process in coastal areas by integrating the recurrent neural network (RNN) model with the analysis of various geomorphological and hydrological features. Its innovative point is to utilize the topographic wetness index (TWI) of every location to classify all inputs into several zones for the model training to ameliorate the accuracy of 2D flood simulations. A 2D numerical flooding model based on hydrodynamic equations was also applied to yield the target inundation depths for the training work of the NN model and adopted to analyze the temporal and spatial variation of coastal flooding under a variety of hydrologic conditions. This study focuses on assessing the applicability of the proposed method that allows for enhancing the model stability and efficiency in forecasting coastal floods caused by extreme rainfall events and storm surges due to the approaching tropical cyclones. The proposed method is promising to be an effective alternative to reinforce the model's stability and computational efficiency for flood forecasting.