Urban vegetation detection using radiometrically calibrated small-footprint full-waveform airborne LiDAR data

Abstract

This paper introduces a new GIS workflow for urban vegetation mapping from high-density (50 pts./m 2) full-waveform airborne LiDAR data, combining the advantages of both raster and point cloud based analysis. Polygon segments derived by edge-based segmentation of the normalized digital surface model are used for classification. A rich set of segment features based on the point cloud and derived from full-waveform attributes is built, serving as input for a decision tree and artificial neural network (ANN) classifier. Exploratory data analysis and detailed investigation of the discriminative power of selected point cloud and full-waveform LiDAR observables indicate a high value of the occurrence of multiple distinct targets in a laser beam (i.e. ‘echo ratio’) for vegetation classification (98% correctness). The radiometric full-waveform observables (e.g. backscattering coefficient) do not suffice as single discriminators with low correctness values using a decision tree classifier (⩽72% correctness) but higher values with ANN (⩽95% correctness). Tests using reduced point densities indicate that the derived segment features and classification accuracies remain relatively stable even up to a reduction factor of 10 (5 pts./m 2). In a representative study area in the City of Vienna/Austria the applicability of the developed object-based GIS workflow is demonstrated. The unique high density full-waveform LiDAR data open a new scale in 3D object characterization but demands for novel joint strategies in object-based raster and 3D point cloud analysis.