Abstract
Accurate estimation of impervious surfaces contributes to a wide range of environmental studies. Among all the imperviousness mapping techniques, satellite remote sensing has become the practical choice, especially for large areas. This dissertation explored the incorporation of intermediate inputs from partially classified remote sensing images in the classification process of impervious surfaces. A hybrid multi-process classification model integrating a priori and a posteriori classifiers was adopted. Intermediate inputs were derived from the a priori classifiers and used to assist the a posteriori classification.
The implementation of the hybrid multi-process classification model on the impervious surface classification using Landsat ETM+ images suggested that the multi-process classification model was superior compared to single-thread classification model in terms of classification accuracy. The incorporation of intermediate inputs as contextual information and supplementary to spectral information has improved the impervious surface classification significantly. Traditional misclassification problems such as separation of impervious surface and soil were successfully tackled through intermediate inputs. Furthermore, the implementation of road structural intermediate inputs demonstrated that exclusive use of intermediate inputs in portions of the image matched or improved classification accuracy obtained solely from spectral information. The usage of intermediate inputs is independent and flexible and the concept of intermediate inputs can be applied on numerous classification tasks and spatial resolutions.