Buildings are one of the most important geospatial features for spatial analysis and mapping. Building extraction has been an active research topic in computer vision as well as digital photogrammetry in recent years. Building detection is the process of obtaining the approximate position and shape of a building, while building extraction can be defined as the problem of precisely determining the building outlines, which is one of the critical problems in digital photogrammetry. Building information is extremely important for many applications such as urban planning, telecommunication, three-dimensional city modeling, or extraction of unauthorized buildings over agricultural lands. Three approaches for building detection based on maximum likelihood classification have been compared, firstly, building detection from
classification of multispectral satellite image only. The second approach is building detection from classification of multispectral satellite image, while the height information from Light Detection and Ranging (LIDAR) data is applied as an additional channel together with spectral channel. The third approach is building detection based on classification of multispectral satellite image where normalized difference vegetation index (NDVI) and the height information from LIDAR data are applied as additional channels together with spectral channel. The contributions of the individual cues used in the classification have been evaluated. The three approaches were tested using urban blocks containing different sizes, roof color and shapes of buildings. The results show that the third approach is the best for building detection followed by the second approach then the first approach. The third approach appears to be quite successful
especially in solving the problem of building detection for those urban blocks that contain closely located buildings as well as in separation of buildings from trees. The third approach results have been improved by developing a building detection module based on integration of classified image, elevation data, and spectral information. A rule-based expert system consists of essentially hypothesis (output; buildings), and variables of a knowledge base were developed in the knowledge engineer of ERDAS Imagine for post classification refinement of initially classified output building mask. Classification rules were enriched with ancillary data such as the normalized digital surface model and the NDVI.
Each rule is a representation of each node in the tree that describes a building class or probability of presence of buildings pixel. Then, the building detection result has been evaluated. It has been found that the use of an expert system, which considers expert knowledge, would further help in the discrimination of the classes and improve classification accuracy of buildings. The overall accuracy of expert classification was 96% and kappa coefficient was 0.95.
|
