@MASTERSTHESIS\{IMM2007-05199,
    author       = "J. P. Andersen",
    title        = "Datafusion and semi-automatic map updating",
    year         = "2007",
    school       = "Informatics and Mathematical Modelling, Technical University of Denmark, {DTU}",
    address      = "Richard Petersens Plads, Building 321, {DK-}2800 Kgs. Lyngby",
    type         = "",
    note         = "Supervised by Assoc. Prof. Allan Aasbjerg Nielsen, {IMM,} {DTU,} external supervisor was Kristian Keller ({COWI} A/{S,} Kgs. Lyngby).",
    url          = "http://www2.compute.dtu.dk/pubdb/pubs/5199-full.html",
    abstract     = "This thesis deals with semi-automatic map updating by the use of fusion of image data and elevation data. Three methods are developed for change detection of buildings in a digital map database. The data used for change detection consists of digitally acquired multispectral aerial imagery, digital elevation data, and the existing map database.

A part of this thesis also deals with the visual comparison of analogue scanned aerial images and digitally acquired aerial images, as well as a comparison based on information content and image noise. This comparison shows a significantly higher level of information content and much lower noise in the digitally acquired images.

The image data for change detection consists of a 4 layer image of {RGB} and nearinfrared (NIR). The elevation data consists of a digital surface model (DSM) and a digital terrain model (DTM) used for the creation of a normalised digital surface model (nDSM). The map database is the TOP10DK from the National Survey and Cadastre - Denmark (KMS).

All three change detection methods are based on the unsupervised classification algorithm, Fuzzy Maximum Likelihood Estimation (FMLE), of either 5 layer images or 4 layer images. The algorithm both takes the spectral characteristics and the spatial characteristics into account. The 5 layer images contains the {RGB,} {NIR,} and the logarithm of the nDSM.

The first method is based on the {FMLE} classification of the 5 layer image. The two clusters that appears to reflect buildings are extracted, merged to one class and compared with the existing buildings in the map database. This method detects the majority (at least 80 \%) of the supposed new buildings. However, several false alarms complicates the change detection.

The second method involves the {FMLE} classification of the 4 layer image and extraction and merging of the building clusters. After that the method is divided into two submethods {2A} and {2B}. In {2A} an intersection between the building class and a Object Above Terrain (OAT) class is performed. The {OAT} class is derived from the nDSM where all elevations are above 2.5 m. The final building class is used for change detection. Method {2B} also involves the class intersection with a {NDVI} class that describes all non-vegetated areas. Both submethods detect the majority (at least 75 \%) of the supposed new buildings with a maximum of 10 false alarms.

In the third method, {FMLE} classification of both 4 and 5 layer images overlaid with a TOP10DK mask is performed resulting in a training sample where only existing buildings are clustered. This training sample is used in a supervised classification algorithm, Maximum Likelihood Estimation (MLE), applied with a Mahalanobis distance threshold. The method is also tested with a prior probability image defined by the existing buildings in the map database. This third method performs worse than the two methods above. The Mahalanobis distance threshold is found to be too low. However, the inclusion of the prior probability image show to have an significant effect.

The conclusion is that the {FMLE} classification algorithm can be used for change detection with certain reservations. It can not be used creation of a training sample for {MLE} classification unless the most optimal Mahalanobis distance parameter is found."
}