本研究主要探討類神經網路結合衛星影像資訊對於颱風雨量預報之影響，並採用自組特徵映射網路(SOM)辨識紅外線與可見光兩種衛星影像，以期從衛星影像中獲得有用的資訊。 以2000～2007年之27場颱風為例，建構石門水庫上游集水區於颱風時期未來一至三小時及六小時累積降雨之六種降雨預報模式：多變量線性迴歸模式(MLR)、倒傳遞類神經網路(BP)、自組特徵映射網路結合倒傳遞類神經網路(SOMBP)、自組特徵映射網路結合多變量線性迴歸(SOMMLR)、SOMBP再結合倒傳遞類神經網路(SOMBPI+BP)以及SOMMLR再結合倒傳遞類神經網路(SOMMLRI+BP)。分為全天及白天兩種不同時段三種方案以衛星、雨量、風速與氣壓等資料組合出七種不同的輸入因子組合。以此三種方案組合探討氣象衛星的資訊對於降雨預報之成效。 預報結果顯示以MLR及BP預測未來累積雨量即可獲得不錯的結果；由拓樸圖可發現SOM具有分辨衛星雲圖特徵的能力，結合雨量資料，可有效改善模式預報值，風速與氣壓資料對於預測效果改善模式幫助不大，顯示衛星影像對於降雨預報上有重要的影響性，可以有效改善模式預報値。 The main purpose of this study is to explore the influence of satellite imagery and meteorological data on typhoon rainfall forecast using artificial neural networks. The self-organizing map (SOM) is adept at recognizing infrared and visible images and can extract some useful information. In this study, six watershed rainfall estimation models are constructed to forecast the amount of rainfall for one, three and six-hour totals during typhoon events. The models are based on SOM, back-propagation neural network (BPNN) or linear regression to investigate the characteristics of satellite imagery information and its influence on rainfall forecast. Twenty-seven typhoon events are collected from 2000 to 2007. The available data are GMS-5/MTSAT remotely sensed data, hourly rainfall data of sixteen rainfall gauge stations of the Shihmen watershed, wind velocity and atmospheric pressure data of three meteorological observation stations. In order to investigate the characteristics and compare the performance among the different models, we design different cases for forecasting the rainfall totals in the daytime and the whole day. Six different models, multivariate linear regression model (MLR), back-propagation neural network (BP), self-organizing map linking with BP (SOMBP), self-organizing map linking with linear regression (SOMMLR), SOMBP linking with BP (SOMBPI+BP) and SOMMLR linking with BP linear regression (SOMMLRI+BP), are constructed to forecast rainfall totals. Seven different combinations of the inputs are used to investigate the effect of rainfall forecast. The results show that (1) the MLR and BP models have nice performances when the input variable only include the past rainfall totals of gauge stations, (2) SOM indeed has the ability to extract patterns from satellite data, (3) SOM can improve results when the rainfall totals are joined, (4) the wind velocity and atmospheric pressure data are helpless for rainfall forecast. The satellite imagery information is indeed helpful to improve the accurate of rainfall forecast.
