Using lidar and aerial photography to build a geographic object database tuned for ecological model

Abstract

This study area covers approximately 16800 square kilometers with a very fragmented landscape. A dataset including aerial photographs at 0.25 cm resolution and LIDAR at 0.8 pts/m has been provided by the Walloon region for the study. The data were resampled at 2m resolution for the purpose of the analysis. The data processing workflow includes three steps : pixel-based image classification, image segmentation and object-based integration. Pixel-based image classification consists in a supervised classification with the spectral values from the aerial photographs (NIR/Red/green/blue), the Digital Height Model extracted from the LIDAR and the intensity of the first LIDAR return. This yielded a classification into broadleaved trees, needleleaved trees, grass, bare soil, crop, pavement, building, water and shadows with more than 80% overal accuracy. The image segmentation approach is the main novelty of this research. In order to fit with the biotopes, image segments indeed had to take the type of slope into account. This was achieved by computing pseudo-hillshades for North-South and West-East orientation and including those two files together with the spectral information from the aerial photographs. The result of this analysis is a set of topographically relevant ecotope delineation. The last step applied contextual decision rules to consistently aggregate the land cover information at the ecotope level and add more information from ancillary datasets.