An Improved Approach for Automatic Detection of Changes in Buildings

Automatic detection of buildings and changes in buildings from airborne laser scanner and image data is discussed. A new, improved method for change detection between an existing building map and building detection results has been developed. Corresponding building objects between the two datasets are found by analysing the overlaps of the buildings. Depending on the correspondences, change detection is carried out, and new, demolished and changed buildings are found. Detection of changed buildings is based on analysing overlap percentages or investigating the building detection results inside and outside buildings on the map by using buffers. Additional rules were developed to investigate tree cover or a digital surface model (DSM) in cases where misclassifications in the building detection stage are likely. The change detection method was evaluated by using suburban test areas covering 4.5 km. Reference results were created by applying the same method to two real building maps. Accuracy estimates for different change classes and building sizes are presented. For all buildings, the completeness and correctness were about 70%. Further tests on building detection with a classification tree based approach are also presented. The method was applied to a new dataset containing laser scanner data and an ortho image created from digital aerial images. Generally, the results are in agreement with our earlier tests and show a mean accuracy of 89% for buildings when compared with a building map, pixel by pixel. Contrary to our earlier study, however, the use of the aerial imagery clearly improved the results. * Corresponding author