The performance of pixel window algorithms in the classification of habitats using VHSR imagery

Abstract

This study investigates the potential of three advanced pixel window classification methods for habitat mapping, namely Kernel based spatial Re-Classification (KRC), Radial Basis Function (RBF) neural networks (NN) and Support Vector Machines (SVM). KRC classifier takes into account the spatial arrangement and frequency of spectral classes present within a predefined square kernel. On the other hand, RBF-NN and SVM classifiers use a set of spectral parameters (digital numbers of training pixels, mean values and standard deviations within a specified window kernel) as input information. The fuzzy means clustering algorithm is utilized for training the RBF networks. This method is based on a fuzzy partition of the input space and requires only a short amount of time to determine both the structure and the parameters of the RBF-NN classifier. The radial basis function is also adopted as the kernel function in the implementation of the SVM methodology. The test area of the present study is Lake Kerkini, a wetland ecosystem located in Macedonia (Northern Greece). The methods are applied to a very high spatial resolution multispectral satellite image acquired by IKONOS-2. The nomenclature used is EUNIS, a detailed hierarchical habitat classification scheme. Several classification experiments are carried out using the same training samples in order to study the behaviour of the three classifiers and perform meaningful comparisons. Overall, all three classifiers performed satisfactorily; however the SVM and RBF-NN classifiers consistently outperformed KRC, reaching overall accuracies of 72% and 69%, respectively.