Skip to main content
Conference paper - Peer-reviewed, 2006

Single tree detection in high resolution satellite images and digital aerial images using artificial neural networks and a geometric-optical forest model

Olofsson, Kenneth; Hagner, Olle


The presentation reports on the work and preliminary results of an on-going 3-year project aimed at the development of advanced methods for detection and measurement of single trees in high-resolution satellite imagery e.g. IKONOS, Quick Bird and airborne optical sensors. The apparent radiance pattern of single tree canopies and corresponding cast shadows depends on the viewing and illumination directions, tree species, topography and atmospheric conditions. There is also a large variation among trees of the same species, due to age, size, branch structure etc. The patterns are also affected by the spatial arrangement of neighbouring trees due to cast shadows and occlusion effects. If the forest is dense and regularly spaced the contrast between the sunlit part of tree crowns and shaded background allows for the use of simple “blob segmentation” algorithms or template matching techniques. However, sparse or open forest conditions motivate more advanced methods that explicitly accounts for the spatial arrangement of neighbouring trees. The detection algorithms developed in this project are based on a learning system approach using artificial neural networks. The networks are trained on datasets generated by a new geometric-optical model developed specifically for this application. Early results show single tree detection, by artificial neural networks, in high resolution remote sensing images using the geometric-optical model


Single tree detection; geometric-optical model; high-resolution satellite images; neural networks; National forest inventory

Published in

Book title: Proceedings: International Workshop, 3D Remote Sensing in Forestry


International Workshop, 3D Remote Sensing in Forestry