洪水預報之精確性取決於上游集水區之雨量資訊是否充足,以往此項訊息皆由地面雨量站觀測所提供,然近十年來,各種遙測影像資訊已相繼應用於降雨之推估,其中氣象衛星影像資料之發展進一步掌握集水區降雨於空間之分佈,其主要優點在於能有效觀察大範圍降雨在時空之變化,故對定量降雨估計而言,衛星影像資訊在某些層面上能提供比地面雨量站更有力的訊息。因此,本研究首先應用美國水文氣象及遙測中心所建立之PERSIANN-CCS系統,藉由衛星影像資訊以推估颱風時期基隆河流載這即時雨量,推估之成果除與雨量站記錄做評比外,並分別架構流量預測模式以評估其優劣,此外,本文更提出一四層架構之加饋式類神經網路,藉由雨量站之觀測與PERSIANN-CCS系統所推估之雨量動用融合方法以獲得雨量融合産品,同時亦就融合雨量之參數及其對流量預測之改善進行探討。
The accuracy of flood forecasting is based on whether the upstream rainfall data is sufficient or not and this information is usually provided by ground rain gauges. Various remotely sensed data have been applied to precipitation estimation over last decade. The development of meteorological satellite sensor further helps to capture the large scale of rainfall distribution. One of the major advantages of satellite imagery is to efficiently detect the temporal and spatial variations of precipitation. As far as quantitative precipitation estimation is concerned, the satellite imagery provides more useful information than ground-based gauges. Therefore, the PERSIANN-CCS built by Center for Hydrometeorology and Remote Sensing (CHRS), UC Irvine, is introduced and used to generate the real-time precipitation over Kee-Lung River during typhoon periods. A comparison of the observed and estimated precipitation and the forecasted hydrologic responses is evaluated. Besides, a four-layer recurrent neural network is developed to produce the merged rainfall products and flood forecasting. Finally, detailed discussions on the merging parameters and the improvement of flood forecasting are given.