Automatic CORINE land cover classification from airborne LIDAR data

Abstract

Point clouds provide valuable information that is not contained in satellite or aerial images. In this work, the potential of airborne LIDAR data for automatic land cover classification following the CORINE standard is evaluated. The methodology consists on the ordering of the point clouds by means of grid maps and rasterized for their use in the training of a Deep Learning classifier model ResNet-50. Three exclusive features of this type of information are extracted: height difference between points, average intensity and number of returns. The methodology has been tested in one case study at level 1 of CORINE inventory, reaching a 73.5% accuracy and a 59,8% Cohen Kappa coefficient. The main confusion occurs between types with strong similarities.