Multisensory Decision Level Fusion for Improvement in Urban Land Classification

Background: Planning and development of urban areas to meet the human developmental requirements is an ongoing process across the globe. This, in turn, requires continuous mapping of urban developmental patterns. Remote sensing and image processing techniques have greatly facilitated the study of urban developmental patterns by mapping urban areas. Objective: The objective of the paper is to carry out the object based classification in urban environments by using fusion techniques on multisensory data to classify natural and man-made objects. Multisensory data fusion using spectral and spatial features is done to improve the classification accuracy. Methods: In our research, the performance of the proposed framework has been verified via investigating the multistage feature level fusion. Spatial and spectral features are explored using feature level fusion between multisensory data and then the database is classified using linear SVM classifier. The individual probabilities (confidence measures) from all such pair of binary SVMs are combined to uniquely represent the object feature to one of the classes. After summing up all the probabilities the class with highest probability value represent the object through decision level fusion. Results: The results explored spatial and spectral features in a unique way using connected component analysis. Also, the improvement in classification accuracy has been achieved. Conclusion: The results reveal that the overall accuracies of SVM classifier are in the range of 53% to 90% for various classes which is improved to 96-98% by majority voting rule.