The OSPARK Algorithm

For the metric-based classification the OSPARK (Optimized Spatial Reclassification Kernel) algorithm has been developed. The algorithm facilitates the reclassification of per-pixel classified remote sensing images, using moving kernel-level (or moving window-level) spatial metrics. The algorithm is based on the SPAtial Reclassification Kernel (SPARK) algorithm, which has been extended to automatically optimize the kernel size (size of the moving window) to the spatial variation detected around the center pixel of the kernel.

In this study the adjacency event spatial metric is calculated for each pixel iteratively for kernel sizes with a radius ranging from 1 to 30 pixels. This metric results for each pixel in a matrix (M). Within each M-matrix the frequency and spatial arrangement of the pixel-based classes positioned next to each other as well as diagonally are counted. Each pair of pixels is called an adjacency event. The results of counting the adjacency events are expressed by an adjacency event matrix.

During each iteration step, i.e. for each kernel size, the adjacency event matrix of each pixel is compared with the template adjacency event matrices (Tk). The Tk matrices are calculated from template kernels that are representative for the land-use classes to be derived and are in that sense comparable to training areas in per-pixel classifications. Since OSPARK iterates over the kernel sizes, the size of the template kernels should match the kernel radius of each iteration step. The comparison of the M matrix with the Tk matrix results in a delta k value, which is the index of similarity.

Delta k can range from 0 to 1. If delta k equals 0, M is completely different from any Tk, while a value of 1 means that they are identical. Finally, after all iterations, each pixel in the input image is assigned to the land-use class of the corresponding Tk matrix with the first local maximum in delta k when increasing the kernel size. The local maximum is chosen, because larger template matrices tend to be less unique for a particular land-use type. Therefore after one or several local maxima delta k will increase with increasing kernel size. The local maximum should be above a user-defined minimum delta k value to prevent classification results with pixels with a low delta k value.

More information about the SPARK algorithm can be found here:

http://pcraster.geo.uu.nl/projects/spark/index.html