臺灣地區易受到颱風暴雨侵襲，常發生水災，且因地形山高河短，淹水過程相當短暫，又受天氣影響，不適合使用飛機或由平流層上方的可見光衛星觀測完整災區，故最適合調查災區淹水區域之手段即為使用可穿透雲層之合成孔徑雷達（SAR）衛星。 本研究使用倒傳遞類神經網路（BPNN）模式及多變量線性迴歸（MLR）模式，結合SAR衛星影像資料，以建構淹水區域辨識模式。其中，BPNN模式可分為僅用淹水時SAR影像之模式一，以及使用淹水前及淹水時兩張SAR影像之模式二，輸入變數包含各像素點經轉換後之雷達散射截面（RCS）值、自身及其鄰近9宮格之統計平均值、標準差、最小值及最大值；MLR模式使用淹水前及淹水時兩張SAR影像，輸入變數兩張影像中各像素點之RCS值差異量、自身及其鄰近9宮格RCS值差異之統計變異數。 結果顯示BPNN模式有較佳的辨識效果，訓練資料與測試資料之淹水辨識正確率分別高達80%與73%以上。錯誤辨識區域大多為分佈零散、未集中於特定區域；為修正這些小而零散的區域，使用型態影像學運算處理將模式輸出結果進行修正，修正後結果正確率大為提升，辨識正確率可提升至90%以上。 Typhoons and storms hit Taiwan several times each year and they cause serious flood disasters. The rivers are short and steep, and their flows are relatively quick with floods lasting only few hours. Due to the factors of the weather, it is not suitable for aircraft or traditional multispectral satellite; hence, the most appropriate way for investigating flood extent is to use Synthetic Aperture Radar (SAR) satellite. In this study, back-propagation neural network (BPNN) model and multivariate linear regression (MLR) model are constructed to identify the flood extent from SAR satellite images. The input variables of the BPNN model are the pixel’s Radar Cross Section (RCS) value and mean, standard deviation, minimum and maximum of RCS values among its adjacent 3×3 pixels. The MLR model uses two images, including the flooding before and the input variables of the MLR model are the difference between the RCS values of two images and the variances among its adjacent 3×3 pixels. The results show that the BPNN model can perform much better than the MLR model. The correct percentages are more than 80% and 73% in training and testing data, respectively. However, the locations of many misidentified areas are very fragmented and unrelated. For correcting the small and fragmented areas, morphological operations are used to modify the outputs of these three identification models. The modified results have been improved a lot and the correct percentages increase up to 90%.
