Improving Classification of Very-High-Resolution Satellite Imagery - Combining Invariant Support Vector Machines and Object-Based Image Analysis to Tackle Limited Information Input

Addressing environmental and socioeconomic challenges in the context of climate change or urbanization, often requires monitoring of large spatial areas. As remote sensing can provide such information, it evolved to be a standard tool to work on related subjects. Image classification often forms the basis for used workflows and derived products. The emergence of new sensor technologies which provide very high spatial and spectral resolution data, made the consideration of objects at finer scales possible and broadened the scope of potential applications of remote sensing. Novel image processing and classification methods such as object-based image analysis and support vector machines, are introduced to effectively exploit the information provided by improved resolutions. Nevertheless, especially for supervised approaches, classification results still depend strongly on the amount and distribution of available ground truth data as information input for training of classifiers. This thesis aims to address the issue by proposing a generic method capable of coping with small ground truth data sets to classify very high spatial resolution data. This is done by transferring invariant support vector machines to the methodology of object-based image analysis. Resulting classifiers appear invariant to scale or geometry representation in ground truth data sets and thus achieve better classification accuracies on limited information input. Experiments on a very-high-resolution image of a complex urban land cover composition - cologne city center - suggest that the proposed method has much scope for future developments. Results show 0.2-0.03 points of kappa accuracy improvement (to reach 0.57-0.79) on small ground truth data sets (20 or less training samples per class) compared to a state of the art classification system for a binary classification problem. Although less pronounced, multi-class settings resemble those tendencies. However, in order to ensure general validity of the results, further research is needed