A New Classification Method for LIDAR Data Based on Unbalanced Support Vector Machine

A building detection method based on multi-spectral image and airborne laser scanning data is presented. Detecting building footprints from laser points cloud data is one of the most difficult problems in building modeling and edge detection. Unbalanced support vector machine can be used to classify laser points from their features, and solve the classification problems of unbalanced dataset with dense buildings and differentiate neighboring tree points from building points. The features include height, height variation information from LIDAR data and spectrum information from registered aerial image. Firstly, candidate building points were selected, and then standard support vector machines were used for preliminary training. The information provided by projected normal difference of training points in normal vector and training size were used to determine the penalty factors of two classes. At last, a new classification hyperplane could be obtained by training again. The experiments show that the classification method based on unbalanced support vector machines can efficiently detect the building points from LIDAR data, and improve the classification accuracy of unbalanced dataset.