An Approximate Spectral Clustering Ensemble for High Spatial Resolution Remote-Sensing Images

Unsupervised clustering of high spatial resolution remote-sensing images plays a significant role in detailed land cover identification, especially for agricultural and environmental monitoring. A recently promising method is approximate spectral clustering (SC) which enables spectral partitioning for large datasets to extract clusters with distinct characteristics without a parametric model. It also facilitates the use of various information types via advanced similarity criteria. However, it requires an empirical selection of a similarity criterion optimal for the corresponding application. To address this challenge, we propose an approximate SC ensemble (ASCE2) which fuses partitionings obtained by different similarity representations. Contrary to existing spectral ensembles for remote-sensing applications, the proposed ASCE2 employs neural gas quantization instead of random sampling, advanced similarity criteria instead of traditional distance-based Gaussian kernel with different decay parameters, and a two-level ensemble. We evaluate the proposed ASCE2 with three measures (accuracy, adjusted Rand index, and normalized mutual information) using five remote-sensing images, two of which are commonly available. We apply the ASCE2 in two applications for agricultural monitoring: 1) land-cover identification to determine orchard fields using a WorldView-2 image (0.5-m spatial resolution) and 2) finding lands in good agricultural condition usingmultitemporal RapidEye images (5-mspatial resolution). Experimental results indicate a significant betterment of the resulting partitionings obtained by the proposed ensemble, with respect to the evaluation measures in these applications.

Kadim Taşdemir, Yaser Moazzen, Isa Yıldırım, "An Approximate Spectral Clustering Ensemble for High Spatial Resolution Remote-Sensing Images", IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, DOI: 10.1109/JSTARS.2015.2424292, 2015.

Approximate spectral clustering with utilized similarity information using geodesic based hybrid distance measures

Spectral clustering has been popular thanks to its ability to extract clusters of varying characteristics without using a parametric model in expense of high computational cost required for eigendecomposition of pairwise similarities. In order to utilize its advantages in large datasets where it is infeasible due to its computational burden, approximate spectral clustering (ASC) methods apply spectral clustering on a reduced set of points (data representatives) selected by sampling or quantization. This two-step approach (i.e. finding the representatives and then clustering them) brings new opportunities for precise similarity definition such as manifold based topological relations, data distribution within the Voronoi polyhedra of the representatives, and their geodesic distance information, which are often ignored in similarity definition for ASC. In this study, we propose geodesic based hybrid similarity criteria which enable the use of different types of information for accurate similarity representation in ASC. Despite the fact that geodesic concept has been widely used in clustering, our contribution is the unique way of representing data topology to form geodesic relations and jointly harnessing various information types including topology, distance and density.The proposed criteria are tested using both sampling (selective sampling) and quantization (neural gas and k-means) approaches.Experiments on artificial datasets, well-known small/medium-size real datasets, and four large datasets (four remote-sensing images), with different types of clusters, show that the proposed geodesic based hybrid similarity criteria outperform traditional similarity criteria in terms of clustering accuracies and several cluster validity indices.

K.Taşdemir, B. Yalçın, İ. Yıldırım, "Approximate spectral clustering with utilized similarity information using geodesic based hybrid distance measures", Pattern Recognition (2014), http://dx.doi.org/10.1016/j.patcog.2014.10.023