Combining Pixel- and Object-based Approaches for Multispectral Image Classification using Dempster-Shafer Theory

We propose an efficient framework for combining pixel and object-based approaches for Remote Sensing Image Classification using Support Vector Machines (SVMs) and Dempster-Shafer Theory of Evidence (DSTE). The pixel-based technique employs the multispectral information for assigning a pixel to a class according to the spectral similarities between the classes, and the object-based technique operates on objects consisting of many homogeneous pixels grouped together in a meaningful way through image segmentation. In order to manage the conflict that results from using both approaches, the final decision is performed using DSTE's rule combination based on probabilistic output from both SVM classifiers. The evaluation test conducted on ETM+ image of Landsat-7 shows that the proposed technique achieved 95.24% classification accuracy. This performance is 5.78% higher than the better accuracy obtained by both SVMs. The proposed combination framework outperforms traditional methods by 2.14% accuracy's margin.