Integration of Feature Selection Techniques for Satellite Image Classification

Feature selection methods have the potential to enhance the classification performance for the sparse datasets with number of features. In this paper, an attempt has been made to assess the potential of these methods over complex and multidimensional remote sensing datasets. The wrapper classes of methods have been adopted for optimal textural feature selection and to analyze the classification results. The most popular feature selection methods in this category are Sequential Feature Selection have been tasted over features extracted from the high resolution satellite images. The sequential feature selection methods are applied to select the best suitable texture feature for landuse classification of the optical datasets. Various textural features per pattern from different texture models such as Gray Level Co-occurrence Matrix (GLCM), Wavelet Transform have been computed for inclusion with SVM based classification. The satellite image has been categorised into 4 land cover classes viz. Water, Unvegitated Island, Vegitated Island and Sand. The accuracy of the classified image has also been assessed using the ground truth data. The manuscript attempts to determine whether the classification error can be reduced by applying feature selection methods to all the features to obtain the optimal set of features. The optimal feature set is used for classification based on their classification scores. The SFFS method has outperformed in terms of optimal feature selection for classification as compared to other feature selection methods.