Geodesic based hybrid similarity criteria for approximate spectral clustering of remote sensing images

Spectral clustering has been successfully used in many applications thanks to its ability to extract clusters with various characteristics without a parametric model and its easy implementation. However, due to its computational cost and memory requirement, it is infeasible for big data such as remote sensing images and it can only be applied through data representatives (obtained by quantization). This approach, approximate spectral clustering (ASC), not only exploits the advantages of spectral clustering for big data, but also enables representing detailed data characteristics in different aspects including topology, local density distribution, Euclidean or geodesic distance. This study presents geodesic based hybrid similarity criteria harnessing different types of information for ASC and shows their performance in extraction of agricultural lands.