Spatiotemporal inferences for use in building detection using series of very-high-resolution space-borne stereo images

Abstract

Automatic building detection from very-high-resolution (VHR) satellite images is a difficult task. The detection accuracy is usually limited by spectral ambiguities and the
\nuncertainties of the available height information. Feature extraction and training sampling collection for supervised methods are other sources of uncertainty. Most widely used VHR sensors have shorter revisit cycles (IKONOS/GeoEye-1/2, 3 days; WorldView 1/2, 1.1 days) due to large off-nadir viewing angles and hence are able to
\nperform consistent acquisition of mono or stereo images. In this article, we investigate the possibility of using high-temporal stereo VHR images to enhance remote-sensing
\nimage interpretation under the context of building detection. Digital surface models, which contain the height information, are generated for each date using semi-global
\nmatching. Pre-classification is performed combining the height and spectral information to obtain an initial building probability map. With a reference land cover map available for one date, the training samples of the other dates are automatically derived using a rule-based validating procedure. A spatiotemporal inference filter is developed
\nconsidering the spectral, spatial, and temporal aspects to enhance the building probability maps. This aims at homogenizing the building probability values of spectrally
\nsimilar pixels in the spatial domain and geometrically similar pixels in the temporal domain, while being robust to the silhouette of the images and geometric discrepancies
\nof the multitemporal data. The effectiveness and robustness of the proposed method are evaluated by performing three experiments on six stereo pairs of the same region over a time period of five years (2006–2011). The area under curve (AUC) of the receiver operating characteristic and kappa statistic (κ) are employed to assess the results. These experiments show that spatiotemporal inference filtering largely improves the accuracy of the building probability map (average AUC = 0.95) while facilitating building extraction in snow-covered images. The resulting building probability maps can be further used for other applications (e.g. building footprint updating).