Adaptive Matched Filters for VHR Image Classification

This paper presents an innovative method for the design and application of matched filters. We extend the use of the matched filters from single object enhancement to multi-object enhancement, and show their potential in improving the pixel-based classification. The challenging step in de-signing the matched filters is the choice of the coefficients, to which we propose a technique based on the use of an optimization algorithm, particle swarm optimization. The classification of the image is performed through the Support Vector Machines. We generated adaptive 2D kernels, used varying sizes and assessed their impact on the classification accuracy. A significant improvement has been achieved using our method compared to the original features alone, which is represented by a gain that ranges between 2.96%-9.43%, and 3.52%-11.7%, for respectively overall and average classification accuracy. Moreover, a very important improvement of the per-class classification accuracy has been achieved, reaching 31.48% gain. Therefore, our extracted spatial features are very discriminative.