SUPPORT VECTOR MACHINES: OPTIMIZATION AND VALIDATION FOR LAND COVER MAPPING USING AERIAL IMAGES AND LIDAR DATA

A method for automatic feature extraction from multispectral aerial images and lidar data based on four study areas with different characteristics is presented. The lidar point clouds were filtered to generate a Digital Terrain Model (DTM) and a Digital Surface Model (DSM) and then the Normalised Digital Surface Model (nDSM) was generated. A total of 22 uncorrelated feature attributes have been generated from the aerial images, the lidar intensity image, DSM and nDSM. Finally, the Support Vector Machines (SVMs) were used to automatically classify buildings, trees, roads and grass from aerial images, lidar data and the generated attributes with the most accurate average classifications of 96.8 % being achieved. Four SVM kernel models (Gaussian Radius Basis Function (RBF); Linear; Polynomial; and Sigmoid) were tested. Also, a new method based on a curve fitting technique was applied to minimize the grid search for the RBF parameters. In addition, the transferability of the RBF parameters from one data set to another was evaluated. Further, we evaluated the contributions of the individual attributes to the quality of the classification results and compared the results against two classifiers of different characteristics (the Self Organizing Map (SOM) and the Regression Trees (RT)). The results demonstrate the greater ability of the SVMs to take advantage of many more attributes compared to other classifiers. The results also showed the relative importance of the nDSM and its attributes for the detection of the different land cover classes using the three classifiers