高解析度衛星影像分類應用於國土利用調查之研究

Abstract

本研究採三幅不同土地覆蓋複雜度SPOT 5衛星融合影像作為實驗區，除了針對不同分類演算法進行成果比較外，並分析加入紋理統計量之前後成果，以及加入空間相關性之不同分類模式，探討影像的高空間解析度對於影像土地覆蓋分類精度是否具有提升效益。 實驗成果以誤差矩陣作展示及精度評估。研究成果顯示，在不同分類演算法之分類成果，得出物件導向分類方法於土地覆蓋複雜之分類成果為佳，且所分類之類別區域較為合理，少有歸屬為其他類別的零碎像元參雜其中。支持向量機在三個實驗區的分類成果非為最佳，主要是因為支持向量機在求取最佳參數組合時，因土地覆蓋類型過於複雜，故無法模擬出最適超平面以區分類別，而使其分類成果較倒傳遞神經網路不佳。 在加入空間相關性之分類成果，顯示不同紋理統計量於同一類別上的貢獻和取樣視窗尺寸十分相關，若取樣視窗過大，則紋理統計量所得成果，差異並不大，但對於紋理特徵相異之類別，紋理統計量的加入帶來的效益便有明顯區分。而且，紋理影像的加入對提升精度效果有限，某些紋理影像加入，甚至降低分類精度，故紋理特徵的選用及其參數的設定必須詳加考量。且加入PCA與加入Rough Set之分類模式不但具有顯著差異，而且透過Rough Set所萃取紋理之線性組合，相較於主成份分析法之萃取，較能有效提昇分類精度。

Three scenes of SPOT-5 high resolution imagery produced with image fusion are utilized for experiments in this study. Several classification schemes based on different methodology are comparatively studied for Land Cover and Land Use applications. The effectiveness of including texture measures in the classification is also analyzed. The experiments are assessed with error matrices and also accuracy indices. It is shown that the classification schemes based on object, which takes the spatial homogeneity into consideration, performs better. Not only the classification accuracy is better, but also less scattered errors are found. Contradicting with earlier studies, Support Vector Machines (SVM) was not found to be better than neural network. This may result from the high complexity of the images under study. In some cases, SVM failed to produce the best hyper plane. Regarding the effectiveness of textures based on GLCM, the sample window size is found to be very critical. Different land cover types are best suited with different window size. In the experiments conducted, introducing texture does not always improve the classification accuracy. The PCA and the method based on Rough Set are compared for the two stages schemes, which utilize a feature selection scheme before the classification. It is found that Rough Set performs better than PCA.