Optimization of matched filters for VHR image segmentation

Abstract

The availability of very high resolution remotely sensed images has made the use of spectral information alone insufficient for class recognition, so the integration of spatial information into the segmentation task becomes necessary. Spatial information takes into account the surrounding of a pixel, rather than dealing with a pixel as an isolated item. In this paper, inspired by matched filtering theory, we propose a new technique for spatial feature extraction. The technique consists of using 2D kernels convolved with the spectral bands. The convolution operation bears valuable spatial information about the pixel under consideration. The concatenation of the extracted spatial features and spectral features of pixels feeds a support vector machine (SVM) to segment the image of interest. To cope with the complexity of images having objects of varying sizes (i.e. urban areas), we adopted a hierarchical strategy. That is, kernels with increasing size are applied to extract different spatial features corresponding to different objects. Thus, we have extended the use of matched filters from single object enhancement to multi-object enhancement. A challenging step in designing the matched filters is the two-dimensional kernels coefficients selection, which we formulated as an optimization problem within a particle swarm optimization framework. The optimization process was driven by two different fitness functions, the cross-validation SVM accuracy and the Bhattacharyya distance, which were both evaluated on training samples. We assessed the proposed procedure on two very high-resolution images having different spatial resolutions. The obtained results showed significant improvement in terms of the overall classification accuracy (over 10% for both images). Moreover, visual inspection of the segmented images, in addition to pepper and salt elimination, revealed significant improvement in many objects not detected by the spectral method. Our method of extracting spatial features seems to be very efficient