Identifying data centers from satellite imagery
Adam Buskirk · Open PRAIRIE (South Dakota State University) · 2016
We develop two different descriptors which can be utilized to describe satellite imagery. The first, the differential-magnitude and radius descriptor, describes a scene by computing the directional gradient of the scene with respect to a vector field whose solutions are circles around a pixel to be described, and then counts pixels in a descriptor matrix according to the magnitude of this gradient and the distance at which this magnitude occurs. The second, the radial Fourier descriptor, extracts from the scene a sequence of annuloid sectors, and uses this to approximate the behavior of the image on a circle around the point to be described. The fast Fourier transform is then used to obtain a description of this function in the frequency domain; the absolute values of these complex-valued frequencies form the descriptor. A set of data to test and perform parameter selection for these procedures using 79 Landsat 8 imagery scenes was constructed. A cellular evolutionary algorithm was then utilized for parameter selection by training and testing support vector machine classifiers using LIBSVM from this dataset utilizing classification accuracy as a fitness function. We then analyze the classification success associated with the two methods equipped with their optimized parameters.