Statistical modelling of urban objects from noisy and low resolution satellite imagery

Kurzfassung

In this article we present a preliminary proof of concept for modeling urban areas using imperfect, high-resolution satellite stereo data - such as oblique Pléiades or WorldView scenes, or older Ikonos and Quickbird imagery with ground sampling distances (GSD) between 50 cm and 1 m. In these stereo scenes, the derived digital surface models (DSMs) are inherently noisy, featuring a vertical resolution of also roughly 50 cm to 1 m. Orthorectifying the satellite imagery against such noisy DSMs produces blurred and distorted orthoimages. To resolve this, our approach introduces statistical methods to extract key urban objects - namely trees, buildings, and roads - from this degraded data. Because the single stereo scene is often acquired during a season when vegetation is difficult to identify in the scene using the Normalized Difference Vegetation Index (NDVI), we first process a Sentinel-2 time series to reliably isolate vegetated regions. Combining the panchromatic stereo DSM, this multi-temporal vegetation mask, and the blurred orthoimage enables an initial spectral and height classification into low-elevation classes (road, soil, vegetation, water) and high-elevation classes (trees, buildings). Tree modeling is then performed statistically by setting up and refining a Voronoi diagram based on local DSM maxima in the tree-classified regions. Buildings are reconstructed by deriving primary directional alignments from the DSM slope and aspect, along with outlines of separable building levels, detected roof ridges, flat roofs, and break lines. For road network extraction, a digital terrain model (DTM) derived from the DSM is combined with a spectral classification of the orthoimage to identify flat road segments and their boundaries, which are afterwards integrated into a statistical model using a network graph. Finally, we manually evaluate the complete pipeline on an available Pléiades scene and compare the results against ground-truth data using random sampling.