Mapping Urban Materials Using High Spatial Resolution Satellite Data

The United Nations expects the world’s urban population to double by 2050 and as cities expand, the replacement of natural land-covers (e.g. soil and vegetation) by impervious surfaces (e.g. concrete, asphalt, and metal) will have significant environmental consequences, including reduction in evapotranspiration, more rapid surface runoff, development of an urban heat island (UHI), and reduction in air and water quality. The monitoring of these urban conditions is necessary to guide sustainable urban growth in a socially and environmentally sensible manner and remote sensing technology may provide a time- and cost-effective means to map and monitor several key variables. This study developed and evaluated an approach for mapping urban materials with Quickbird satellite data and lidar using pixel-based and object-based image classification techniques on a subset of Brisbane, in Queensland, Australia. Most urban mapping studies using multi-spectral imagery have mapped cities at broad categories of land use/land-cover but this study selected map classes at a more detailed level. The mapping level focussed on the materials that make up the surfaces in urban environments for which there are known emissivity values, which could be used to map urban energy exchanges. This created the potential to utilise the results from this work to model surface energy fluxes, and consequently, land surface temperature in a subsequent study. The ability of the local government in Brisbane to utilise a fast and economical approach to obtaining temperature information at a fine spatial scale, i.e. with pixels < 10 m, would help to improve city planning decisions and mitigate the adverse effects of urbanisation on the environment. Quickbird imagery, which has proven useful for urban mapping applications because of its high-spatial resolution, was used in this study. Lidar data, which have also been successfully used in urban mapping, were incorporated in the classification process with the Quickbird imagery. The classification methods of this study began with pixel-based classification, then object-based classification, and finished with an integration of the two. Each classification step played a key role in identifying the best data, method, and scale parameters (for object-based methods) for use in the final step of the methodology. First, the pixel-based classification determined a suitable level of detail for map classes and identified which additional image bands were best for discriminating the selected classes. This method had high accuracy but a great deal of the salt and pepper effect, with many real-world image objects (e.g. roads and building roofs) misclassified as a mix of various materials. The misclassifications of this method were attributed to spectral overlap between classes and lack of strong normal distributions for the classes within the individual image bands. Two object-based classifications determined the appropriate scale parameters for image segmentation and the best object features for identifying the urban materials classes. The standard nearest neighbour classification had the lowest overall accuracy and poor delineation of real-world objects, particularly within dense residential areas. This resulted from use of only one segmentation level and one set of object features for the classification of all image objects. The second object-based classification utilised a stepwise method that classified image objects in a hierarchical manner. This method showed improvement from the previous classifications and successfully delineated most of the urban materials; improvement was attributed to the use of multiple object levels and unique object features in each class description. Error assessments, conducted using standard error matrices for all classifications, identified strengths and weaknesses of each mapping approach. Across all methods, natural materials (e.g. water and trees) as a whole had the highest classification accuracies, while built materials (e.g. concrete and metal) had the lowest. The pixel-based classification had the highest overall accuracy and the standard nearest neighbour classification had the lowest. The stepwise classification achieved comparable accuracy to the pixel-based classification but with less salt and pepper and improved identification of real-world features. Thus, the final approach developed for mapping urban materials integrated pixel-based and object-based classification methods which allowed spectral values as well as colour/shape/class-related features to be utilised on multiple object levels, with unique class descriptions for each urban material. The integrated classification had high overall accuracy and provided the best delineation of real-world image objects. Though the developed mapping approach was only tested on a small subset of Brisbane, this area contained a good representation of the materials present within urban environments and also utilised band ratios, thereby improving the ability for these rule-sets to be applied to other urban regions. Follow-up studies will determine the actual success of the mapping approach on additional sites and the suitability of the results from this work to be utilised for energy flux modelling