Classification of terrain objects using hyper-dimensional (multi-temporal multi-spectral) images through purpose-oriented feature extraction

Multi-temporal multi-spectral observations produce huge amounts of data, especially when hyperspectral sensors are used and/or when the number of observations increases. The authors have already developed a method of purpose-oriented feature extraction and successfully applied it to hyperspectral data which has several hundreds of dimensions. The authors apply the basic idea of this method to the multi-temporal multi-spectral remotely sensed images for the classification of terrain objects. A small number of features were extracted according to the purpose or the intention of classification. The features matched to the intention to discriminate some important classes of terrain objects are extracted. The method was applied to the multi-temporal Landsat TM images with success. Results of feature extraction and classification are shown