3D change detection from high and very high resolution satellite stereo imagery

Change detection is one of the most essential processing steps for monitoring urban and forest areas using remote sensing data. Even though 2D data obtained from satellite images from different dates can already provide plenty of useful information, it is usually insufficient when dealing with changes in the vertical direction. Moreover, if only one class of changes, such as buildings or forest, is of interest, it is often difficult to distinguish between relevant and irrelevant changes. In such cases, the information provided by Digital Surface Models (DSMs) is crucial, as it provides additional height information, which can be indispensable when analyzing changes. This dissertation addresses the challenge of using DSMs generated by satellite stereo images for 3D change detection. DSM generation techniques based on stereo imagery from space have been improved continuously in recent years, enhancing the quality of the generated DSMs considerably. Nevertheless, up to now these DSMs have not been widely adopted for change detection methods. Available 3D change detection approaches prefer LiDAR data, which are more accurate but have the drawback of being more expensive and exhibit a comparatively low temporal repetition rate. The characteristic and quality of DSMs based on satellite stereo imagery have so far hardly been considered within 3D change detection procedures. Therefore, more in-depth investigations concerning the adoption of these DSMs for 3D change detection should be performed. In this dissertation, DSMs based on stereo imagery have been visually and numerically evaluated and subsequently analyzed for various land cover types. Taking into account the quality of DSMs generated with the described methods, three DSM-assisted change detection approaches are developed. The first method, called “DSM-assisted change localization”, describes a robust change difference map generation method followed by DSM denoising. The generated change map is refined using vegetation and shadow masks and finally shape feature are used to consolidate the results through distinguishing relevant from irrelevant objects. Concerning fusion-based change detection, two methods, feature fusion and decision fusion, are proposed. The proposed feature fusion methods make use of the fact that panchromatic images feature much sharper contours than DSMs. To alleviate the shortcomings of the DSMs, the designed region-based change detection framework extracts the original regions from the ortho- II images. For this approach, a new robust region-merging strategy is proposed to combine segmentation maps from two dates. Regarding the uncertain information contained in the DSMs and spectral images, a decision fusion method is proposed as the second fusion-based change detection method. The extracted features are classified as change indicators and no-change indicators, while two steps of the Dempster-Shafer fusion model are implemented for the final change detection. Post-classification is a common DSM-assisted change detection method, since the DSM can be very helpful for building extraction. In this third approach, the changed building’s location is obtained by comparing the new building mask with existing (often outdated) building footprint information, e.g. from GIS databases. To extract the boundaries of newly built buildings, a robust building extraction method has been developed by also considering the quality of the DSM. These three approaches are evaluated experimentally using four representative data sets. Quantitative and qualitative experimental results obtained from each data set are analyzed in detail. The experimental results show that all of the proposed approaches are able to determine the change status of the objects of interest. The results achieved vary according to the DSM quality, the object of interest and the test area. Furthermore, it is shown experimentally that, by making proper use of DSMs in complex decision frameworks, both the efficiency and the accuracy of the change detection are improve in comparison to 2D change detection approaches. In addition, the developed approaches enable the rapid localization of changes concerning objects of interest, such as buildings and forest, which is valuable for many applications such as fast response systems after earthquake or other disasters.