影像應用時最爲重要的研究議題就是影像分類。基本上人類分類影像程序是將影像中區塊特徵(Region Feature)萃取出來，合併其特徵後進而形成有效的識別物件。然而，目前機器分類影像時是以逐像元概念(Pixel-based)進行特徵萃取之工作，兩者程序並不相同。爲解決上述問題本研究提出以區塊爲主(Region-based)的區塊化物件分類(Regional Object Classification, ROC)法，這是一個混合非監督程序（區域成長法）與監督式程序（最大概似分類法）的新概念，這個概念可解決逐像元分類法應用於高解析度影像時所造成之嚴重的椒鹽效應(Salt and Pepper Effect)。本研究在雲林縣農業區內選擇高解析度數值航照影像(ADS-40)，針對水稻這個類別作爲實證項目，以逐像元與區域化分類概念進行問題比較。研究結果顯示，透過區塊化物件分類模式可萃取出完整度極高的水稻坵塊田，分類精度可從92.6%(Pixel-based)提高到95.3%(Region-based)。
The image classification is the most important issue in Remote Sensing. In essence, extraction of regional feature is an important process on human’s image classification process to form their effective identification of the target objects. However, vision cognition of machine learning follows Pixel-based which disobeys the rule of human’s image classification process. To resolve the forgoing problems, this study proposes a Region-based approach for Regional Object Classification (ROC). This method is a hyper module which applies an unsupervised method (Region-Growing, RG) to a supervised approach (Maximum Likelihood Classifier, MLC). New concept can effectively reduces a Salt and Pepper effect of classification result from very high resolution image by conventional pixel base classifier. Accordingly, a module of Regional concept is applied to improve its classification accuracy. Our outcomes are used by mean of the ADS-40 data to present the classification performance. Finally, this study utilized the ROC classifier and the overall accuracy of region based concepts (95.3%) is better than those of the pixel-based concepts (92.6%) on the evaluation of paddy rice from ADS-40 image. This result shows that an appropriate progress through regional module can effectively improve the accuracy of image classification.
