Abstract:
Unsupervised classification is often used to process large datasets such as hyperspectral images for which training data (“ground truth”) is difficult to obtain. Recent m...Show MoreMetadata
Abstract:
Unsupervised classification is often used to process large datasets such as hyperspectral images for which training data (“ground truth”) is difficult to obtain. Recent methods based on density estimation with the k nearest neighbours principle are effective to handle this type of data and have the advantage of relying on one parameter only: k, the number of neighbours. However, the classification results are very sensitive to this parameter. In this paper, we propose a new approach to improve the search for relevant nearest neighbours. The method is based on the assumption that clusters are separated by regions of low density in the space spanned by the hyperspectral bands. For a given point, if it is separated from one of its neighbours by a region of low density, then the neighbour is consider invalid. This leads to an increased stability of clustering accuracy with respect to k, especially for convex clusters. In addition, we propose effective methods to automatically estimate the optimal value of k as well as the parameters for local density estimation, thus making the proposed approach non-parametric in addition to being unsupervised, non-iterative and deterministic. Results on synthetic and hyperspectral datasets indicate that our method systematically improves clustering accuracy.
Date of Conference: 04-06 December 2017
Date Added to IEEE Xplore: 05 July 2018
ISBN Information:
Electronic ISSN: 2151-2205