A wavelet-based approach to the classification of remotely-sensed images: A comparison of different feature sets in an urban environment

Classification of remotely-sensed images is a process of automatically categorizing the pixels into land-use and land-cover classes. Many classification approaches have been developed and applied to the remotely-sensed images, and all of them have achieved a certain extent of success. The wavelet-based approach has been proven by many researchers to be an effective way to classify urban classes using multiple scales. Previous research on wavelet-based approaches focused on the effectiveness of the use of different wavelet feature measures; however, no investigation has been conducted to examine the effectiveness of combining other feature sets with wavelet-based features.; In this thesis, we conduct a series of experiments to investigate the effectiveness of various combinations of the different types of feature sets, including a spectral-based feature set, a variance-based feature set and a wavelet-based feature set. All the experiments use the identical study area, training data, reference data, testing data, and classification algorithm while varying feature sets. The classification accuracy from different feature sets is evaluated using the traditional accuracy assessment method from reference data, as well as the confidence levels represented by Inverse-Minimum-Distance (IMD maps) for all pixels. The experimental results show that an IMD map can be used to evaluate the relative pixel labelling confidence for classes within a classified map, but it is not convincing to compare IMD maps generated by different feature sets.; The spectral-based feature set has the basic discrimination power to distinguish classes with middle and high homogeneity values. However, it has little success to correctly classify classes with low homogeneity values, such as the Residential class. Compared with spectral-based feature set, the multi-scale wavelet-based feature set can improve the discrimination power for classes with both low and high homogeneity values. The variance-based feature set alone has little discrimination power, no matter what homogeneity level the class possesses. However, adding the variance-based features into the spectral-based feature set or wavelet-based feature set can dramatically increase the classification accuracy for classes with low homogeneity values.