圖書典藏及影音數位平台

本署已陸續完成各地區製作大量且高精度地形數值之淹水潛勢資料製作與更新，並以空間不均勻分佈之不同降雨重現時距組合產製龐大之淹水潛勢歷程資料，為延伸淹水潛勢資料之應用範圍，以提升淹水潛勢資料在災害防救計畫之參考價值。去年度完成辦理「區域型淹水警示技術研發應用-以宜蘭縣為例」計畫，已建置完成宜蘭縣全區預測未來1～3小時淹水深及淹水分級範圍之預警系統，並於99年10月的梅姬颱風進行模式測試，可即時預警該場颱風造成宜蘭縣各鄉鎮市的淹水趨勢，確實增進淹水潛勢資料之加值應用。爰此，本年度計畫擴展至基隆市、台北縣市、桃園縣等北部區域之區域型淹水預警與推估模式，並滿足區域型淹水之線上即時預測，提供洪災防救應變之重要參考。

With high resolution requirements, the simulations of flood inundation extent need a huge amount of data and computing time by using a series of numerical models. It is really difficult to provide an on-line simulation and forecast of the inundation extent. For extending the applications of the flood inundation maps, this project built regional flood inundation forecasting systems in four northern counties and provided an on-line forecasting system during the period of typhoons and storms. Besides, this project integrated the artificial neural networks, data base, geographic information system, and web interface to establish an on-line regional flood inundation warning system and display the flood depths of villages and towns of these four counties on Google Maps and Google Earth. This project established four regional flood inundation forecasting systems in Keelung City, Taipei City, New Taipei City and Taoyuan County. The modeling process is divided into the data preprocessing and model building stages. In data preprocessing stage, the flooding grids, whose flood depths were higher than a threshold, were selected; then, before clustering analysis, the abnormal or unreasonable grids were deleted or amended. K-means clustering was used to cluster the flooding grids into a number of groups, then, further divide into a number of sub-groups based on the geographic relation and flooded characteristics, and select the control point and representative point for each sub-group. These four counties were divided into 8, 18, 33 and 21 sub-groups, respectively. In the model building stage, the back-propagation neural networks (BPNNs) were built for control points, the linear regression models for the grids that have highly linear correlation with the control point, and multi-grid BPNNs for the grids that do not have highly linear correlation with the control point. The networks are trained and validated by using the inundation maps generated from the designed rainfall patterns. The proposed regional inundation forecast system would only take 2 minutes to make one- to three-hours flood forecast for a 24-hour rainfall pattern, while the conventional two-dimensional numerical model would take over 10 hours to generate 24-hour hydrograph of flood depths. The efficiency has been greatly improved and the proposed system is ready for on-line forecast in the whole region of four northern counties. The on-line regional inundation forecast systems are easily executed through the database to access the data sets and to update the parameters and information obtained from model outputs simultaneously.