Information fusion of aerial images and LIDAR data in urban areas: vector-stacking, re-classification and post-processing approaches

This research investigates information fusion approaches of high-resolution aerial images and elevation data from Light Detection and Ranging (LIDAR) for urban-environment mapping. Three feature fusion methods are proposed and compared: (1) the vector-stacking approach that combines spectral and LIDAR features in one classifier; (2) the re-classification approach that firstly processes spectral signals in a classifier and then integrates its output with LIDAR features to obtain the final result and (3) the post-processing approach that uses the LIDAR data to refine the results of spectral classification. The height features used in the above three algorithms are extracted from the LIDAR digital surface model (DSM) image; these include elevation difference, maximum and minimum values, variance and the grey-level co-occurrence matrix (GLCM) textures. In addition, the average height from object-based segmentation is also computed. In the experiments, support vector machines (SVMs) are used as classifiers for all fusion schemes due to their capability and robustness for many classification problems. The three algorithms are evaluated using a 40-cm spatial resolution digital orthophoto and the corresponding LIDAR data of Odense, Denmark. In the experiments, the vector-stacking method with the Maximum-Minimum (Max-Min) feature, the reclassification method with the Max-Min feature and the post-processing approach obtain promising results (94.7%, 95.0% and 94.6%, respectively), which are significantly higher than the spectral-only classification (82.5%).