| Summary: | Recently, especially within the last two decades, the demand for DSMs (Digital Surface Models) and 3D city models has increased dramatically. This has arisen due to the emergence of new applications beyond construction or analysis and consequently to a focus on accuracy and the cost. This thesis addresses two linked subjects: first improving the quality of the DSM by merging different source DSMs using a Bayesian approach; and second, extracting building footprints using approaches, including Bayesian approaches, and producing 3D models. Regarding the first topic, a probabilistic model has been generated based on the Bayesian approach in order to merge different source DSMs from different sensors. The Bayesian approach is specified to be ideal in the case when the data is limited and this can consequently be compensated by introducing the a priori. The implemented prior is based on the hypothesis that the building roof outlines are specified to be smooth, for that reason local entropy has been implemented in order to infer the a priori data. In addition to the a priori estimation, the quality of the DSMs is obtained by using field checkpoints from differential GNSS. The validation results have shown that the model was successfully able to improve the quality of the DSMs and improving some characteristics such as the roof surfaces, which consequently led to better representations. In addition to that, the developed model has been compared with the Maximum Likelihood model which showed similar quantitative statistical results and better qualitative results. Perhaps it is worth mentioning that, although the DSMs used in the merging have been produced using satellite images, the model can be applied on any type of DSM. The second topic is building footprint extraction based on using satellite imagery. An efficient flow-line for automatic building footprint extraction and 3D model construction, from both stereo panchromatic and multispectral satellite imagery was developed. This flow-line has been applied in an area of different building types, with both hipped and sloped roofs. The flow line consisted of multi stages. First, data preparation, digital orthoimagery and DSMs are created from WorldView-1. Pleiades imagery is used to create a vegetation mask. The orthoimagery then undergoes binary classification into ‘foreground’ (including buildings, shadows, open-water, roads and trees) and ‘background’ (including grass, bare soil, and clay). From the foreground class, shadows and open water are removed after creating a shadow mask by thresholding the same orthoimagery. Likewise roads have been removed, for the time being, after interactively creating a mask using the orthoimagery. NDVI processing of the Pleiades imagery has been used to create a mask for removing the trees. An ‘edge map’ is produced using Canny edge detection to define the exact building boundary outlines, from enhanced orthoimagery. A normalised digital surface model (nDSM) is produced from the original DSM using smoothing and subtracting techniques. Second, start Building Detection and Extraction. Buildings can be detected, in part, in the nDSM as isolated relatively elevated ‘blobs’. These nDSM ‘blobs’ are uniquely labelled to identify rudimentary buildings. Each ‘blob’ is paired with its corresponding ‘foreground’ area from the orthoimagery. Each ‘foreground’ area is used as an initial building boundary, which is then vectorised and simplified. Some unnecessary details in the ‘edge map’, particularly on the roofs of the buildings can be removed using mathematical morphology. Some building edges are not detected in the ‘edge map’ due to low contrast in some parts of the orthoimagery. The ‘edge map’ is subsequently further improved also using mathematical morphology, leading to the ‘modified edge map’. Finally, A Bayesian approach is used to find the most probable coordinates of the building footprints, based on the ‘modified edge map’. The proposal that is made for the footprint a priori data is based on the creating a PDF which assumes that the probable footprint angle at the corner is 90o and along the edge is 180o, with a less probable value given to the other angles such as 45o and 135o. The 3D model is constructed by extracting the elevation of the buildings from the DSM and combining it with the regularized building boundary. Validation, both quantitatively and qualitatively has shown that the developed process and associated algorithms have successfully been able to extract building footprints and create 3D models.
|
|---|
