Cross Track Satellite Stereo For 3D Modelling of Urban Areas

Stereo imagery from very high resolution satellites, such as Pleiades or the WorldView Series can be utilized to create high resolution digital surface models (DSMs). Especially when very high quality tri-stereo acquisitions with small convergence angles are available, buildings can be reconstructed well even in dense city centres. Especially for rapid mapping or crisis-response applications high resolution DSMs would are helpful for tasks such city modelling and ultimately population estimation. Unfortunately, such high quality stereo data is usually not available in the archives, and thus needs to newly acquired in most cases, leading to long delays and high cost. In this work, we utilize existing mono images from the catalogues to produce digital surface models. For such data, temporal changes and non-ideal stereo geometry lead to varying appearance of objects in the images and thus incomplete and noisy digital surface models. In the frame of the X3D4Pop Project, we aim to clarify the state-of-the-art of rapid DSM generation in crisis situations, allowing first responders and humanitarian organisations to gain a better overview of the population affected. We therefore strive to (1) define a set of test scenarios realistically simulating a real crisis/rapid mapping situation, and to evaluate the project outcomes with respect to these scenarios, against reference data; (2) produce a set of DSMs/DTMs (and ortho-images) from various image pairings of different sensors to test the sensitivity of the DEM production chain on differences between the two acquisitions, with the ultimate goal of population estimations using DSMs of different quality dependent of data availability under the defined test scenarios. For this task Pleiades VHR Data over Salzburg and Port au Prince was selected, including both high quality stereo and tri-stereo acquisitions, as well as singe image captures. Dense stereo matching using the Semi-Global Matching algorithms was applied to all combinations of 15 Pleiades images over Port au Prince and 9 images over Salzburg. Results from the digital surface model generation process are evaluated and compared with respect to density as well as accuracy with respect to LIDAR reference data and the suitability for further derivation of digital terrain models and further parameters such as building footprints and heights was investigated. Results show a large difference in density and general quality between the dedicated stereo acquisitions and the multi-temporal datasets. The derivation of terrain models difficult as many temporal changes due to vegetation or road-usage lead to a strongly reduced number of ground points. However, building roofs tend to be more stable and are slightly better represented than ground points in the multi-temporal DSMs. However, this information cannot be extracted using the traditional process of generating digital terrain models and subtracting those from the DSMs to obtain building footprints and heights