Mapping individual tree location, height and species in broadleaved deciduous forest using airborne LIDAR and multi‐spectral remotely sensed data

Abstract

Automated feature extraction based on prototypes is only partially successful when applied to remotely sensed imagery of natural scenes due to the complexity and unpredictability of the shape and geometry of natural features. Here, a new method is developed for extracting the locations of treetops by applying GIS (Geographical Information System) overlay techniques and morphological functions to high spatial resolution airborne imagery. This method is based on the geometrical and spatial properties of tree crowns. Airborne data of the study site in the New Forest, UK included colour aerial photographs, LIDAR (Light Detection And Ranging) and ATM (Airborne Thematic Mapper) imagery. A DEM (Digital Elevation Model) was generated from LIDAR data and then subtracted from the original LIDAR image to create a Canopy Height Model (CHM). A set of procedures using image contouring and the manipulation of the resulting polygons was implemented to extract treetops from the aerial photographs and the CHM. Criteria were developed and threshold values were set using a supervised approach for the acceptance or rejection of features based on field knowledge. Tree species were mapped by classifying the ATM data and these data were co‐registered with the treetop layer. For broadleaved deciduous plantations the success of treetop extraction using aerial photographs was 91%, but was much lower using LIDAR data. For semi‐natural forests, the LIDAR produced better results than the aerial photographs with a success of 80%, which was considered high, given the complexity of these uneven aged stands. The methodology presented here is easy to apply as it is implemented within a GIS and the final product is an accurate map with information about the location, height and species of each tree.

Acknowledgments

We are grateful to the UK Environment Agency for the provision of LIDAR data and to UK NERC for the provision of colour aerial photographs and ATM data. Many thanks to K. McVay for the script (View.R2Vpoly.ave), which converts raster to polygon. This script gives the user the option to choose whether or not to use generalization (weeding) with the Douglas‐Poiker algorithm. The default script in ArcView performs the conversion only with the weeding option and was not suitable for the purposes of this research. Our thanks to Tara Montgomery for the script which extracts the centroid points from a polygon layer (script name: PolyCentroidToPoint.ave), and to Marco Boeringa from Amsterdam Water Supply for the Kriging extension (Kriging Interpolator 3.1 (extension name: gwa_SAkriginginterpolator.avx). All scripts and extensions were available from www.esri.com/arcscripts.