Classification of Settlements in Satellite Images Using Holistic Feature Extraction

This paper presents a holistic feature extraction technique for the classification of settlements in high resolution satellite images. The goal is to design a system which automatically classifies settlements in large images. In this work Principal Component Analysis (PCA) is used to extract the feature or signature of settlements. These features can be used by different classifiers. Euclidean norm is used to classify the two classes using the features calculated by PCA. A moving window is used for larger images to resample the testing images. In this work 80×80, and 40×40 pixel windows are used for resample as the training images used in this work has these dimension. The accuracy of the system is checked by comparing the actual results with the reference map. The Comparisons is also made between 40×40 and 80×80 dimensions. The settlements are 100% classified with the threshold of 1500 distance measure for 80×80 dimension whereas threshold used for 40×40 is 900. The overall accuracy of 80×80 window is 96.43 where the accuracy of 40×40 window is 89.64. The 80×80 window is good window analyzed in this work, because training sample contains more principal components. The accuracy measured by calculating settlements, non settlements and mix settlements.