K-means algorithm based on stochastic distances for polarimetric synthetic aperture radar image classification

• Autor(a) principal: Negri, Rogério Galante [UNESP]
• Data de Publicação: 2016
• Outros Autores: Da Silva, Wagner Barreto, Mendes, Tatiana Sussel Gonçalves [UNESP]
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1117/1.JRS.10.045005; http://hdl.handle.net/11449/220713
• topic: K-means; polarimetry; stochastic distances; unsupervised classification

The availability of polarimetric synthetic aperture radar (PolSAR) images has increased, and consequently, the classification of such images has received immense attention. Among different classification methods in the literature, it is possible to distinguish them according to learning paradigm and approach. Unsupervised methods have as advantage the independence of labeled data for training. Regarding the approach, image classification can be performed based on its individual pixels or on previously identified regions in the image. Previous studies verified that the region-based classification of PolSAR images using stochastic distances can produce better results in comparison with the pixel-based. Faced with the independence of training data by unsupervised methods and the potential of the region-based approach with stochastic distances, this study proposes a version of the unsupervised K-means algorithm for PolSAR region-based classification based on stochastic distances. The Bhattacharyya stochastic distance between Wishart distributions was adopted to measure the dissimilarity among regions of the PolSAR image. Additionally, a measure was proposed to compare unsupervised classification results. Two case studies that consider real and simulated images were conducted, and the results showed that the proposed version of K-means achieves higher accuracy values in comparison with the classic version.