Object-based interpretation methods for mapping built-up areas

There is a growing demand for high-quality spatial data and for efficient methods of updating spatial databases. In the present study, automated object-based interpretation methods were developed and tested for coarse land use mapping, detailed land cover and building mapping, and change detection of buildings. Various modern remotely sensed datasets were used in the study. An automatic classification tree method was applied to building detection and land cover classification to automate the development of classification rules. A combination of a permanent land cover classification test field and the classification tree method was suggested and tested to allow rapid analysis and comparison of new datasets. The classification and change detection results were compared with up-to-date map data or reference points to evaluate their quality. The combined use of airborne laser scanner data and digital aerial imagery gave promising results considering topographic mapping. In automated building detection using laser scanner and aerial image data, 96% of all buildings larger than 60 m2 were correctly detected. This accuracy level (96%) is compatible with operational quality requirements. In automated change detection, about 80% of all reference buildings were correctly classified. The overall accuracy of a land cover classification into buildings, trees, vegetated ground and non-vegetated ground using laser scanner and aerial image data was 97% compared with reference points. When aerial image data alone were used, the accuracy was 74%. A comparison between first pulse and last pulse laser scanner data in building detection was also carried out. The comparison showed that the use of last pulse data instead of first pulse data can improve the building detection results. The results yielded by automated interpretation methods could be helpful in the manual updating process of a topographic database. The results could also be used as the basis for further automated processing steps to delineate and reconstruct objects. The synthetic aperture radar (SAR) and optical satellite image data used in the study have their main potential in land cover monitoring applications. The coarse land use classification of a multitemporal interferometric SAR dataset into built-up areas, forests and open areas lead to an overall accuracy of 97% when compared with reference points. This dataset also appeared to be promising for classifying built-up areas into subclasses according to building density. Important topics for further research include more advanced interpretation methods, new and multitemporal datasets, optimal combinations of the datasets, and wider sets of objects and classes. From the practical point of view, work is needed in fitting automated interpretation methods in operational mapping processes and in further testing of the methods.