Gabor Wavelets and Gaussian Models to Separate Ground and Non-ground for Airborne Scanned LIDAR Data

To separate non-ground and ground is the first step of terrain feature classification for remotely sensed LIDAR (LIght Detection And Ranging) data. The LIDAR data, which were captured by airborne laser scanner, contain terrain surface height information. This paper aims to segment non-ground (buildings and high vegetation) from ground (flat field) in hilly urban areas. Gabor wavelets were applied to LIDAR data tiles to generate Gabor wavelet features. Training sets of ground and non-ground areas were manually selected from windowed features to establish Gaussian models for the two types of region. In the testing, each window in the Gabor feature space was tested by the probability density function (PDF) of the two models. Maximum likelihood was used in the final classification. The test results have shown that ground areas (flat field) can successfully be segmented from LIDAR data. Most buildings and high vegetation can be detected. In addition, Gabor wavelet transform can remove hill or slope effects on the regular grid LIDAR data by tuning Gabor parameters.