Ensemble PHOG and SIFT Features Extraction Techniques to Classify High Resolution Satellite Images

Abstract

The task of indentifying similar objects within the querying image remains challenging. It is due to viewpoint and lighting changes, deformation and partial occlusions that may exist across different examples. In this framework we focus on combination methods that ensemble multiple descriptors at multiple spatial resolution levels. Ensemble PHOG (Pyramid histogram orientation and gradient) and SIFT (Scale invariant feature transformation) descriptors are used for the feature extraction to achieve the good classification accuracy. Within a region local feature was captured by the distribution over edge orientation, and spatial layout by tiling the image into regions at multiple resolutions. The SIFT features are extracted for each PHOG block. These features are trained using SOM network. Later SVM and Neural network classifiers are used for classification. Results demonstrating the effectiveness of the proposed technique are provided using confusion matrix, transition matrix and other accuracy measures. Area of different land cover regions are calculated, which can be used for land use changes.