Classification of Landsat 8 Imagery Using Kohonen’s Self Organizing Maps and Learning Vector Quantization

Abstract

Image classification is the most commonly employed technique for extricating land cover report from remotely sensed images. In the last two decades, advanced image classifiers have been extensively applied in remotely sensed (RS) image classification studies. Kohonen’s Self Organizing Maps (SOM) and its supervised version Learning Vector Quantization (LVQ) algorithms have been applied in a variety of machine learning and pattern recognition studies as they are derived from neural networks. In this paper, the classification ability of SOM, LVQ-1 and LVQ-2 techniques are investigated using medium resolution Landsat 8 OLI/TIRS RS imagery. The study presents a detailed analysis of SOM and LVQ algorithms in RS data classification based on the LULC separability. The study uses Divergence (Div) and Transformed Divergence (TD) as metrics for measuring LULC separability. Results of SOM and LVQs were compared with conventional maximum likelihood and minimum-distance- to-means classifiers. Supervised Kohonen’s LVQ algorithms produced better accuracies than conventional classifiers. Also, LVQ algorithms showed great efficiency in avoiding overfitting of the dominant classes and separating spectrally overlapping classes. The study is indication of the applicability of Kohonen’s algorithms for RS data classification.