Classification Tree Based Building Detection from Laser Scanner and Aerial Image Data

A classification tree based approach for building detection was tested. A digital surface model (DSM) derived from last pulse laser scanner data was first segmented and the segments were classified into classes ‘ground’ and ‘building or tree’ on the basis of preclassified laser points. ‘Building and tree’ segments were further classified into buildings and trees by using the classification tree method. Four classification tests were carried out using different combinations of 44 input attributes. The attributes were derived from the last pulse DSM, first pulse DSM and an aerial colour ortho image. In addition, shape attributes calculated for the segments were used. The attributes of training segments were presented as input data for the classification tree method, which constructed automatically a classification tree for each test. The trees were then applied to classification of a separate test area. Compared with a building map, a mean accuracy of almost 90% was achieved for buildings in each test. The classification tree method appeared to be a feasible and highly automatic approach for distinguishing buildings from trees. If new data sources become available in the future, they can be easily included in the classification process. The results also suggest that satisfactory building detection results can be obtained with different combinations of input data sources. By using a statistical method, it is possible to find useful attributes and classification rules in different cases. The use of an aerial image or both first pulse and last pulse laser scanner data does not necessarily improve the results significantly, compared with a classification that uses only last pulse laser scanner data. * Corresponding author.