Mapping tree health using airborne full-waveform laser scans and hyperspectral imagery: a case study for floodplain eucalypt forest

Friday, 19 December 2014
Iurii Shendryk, Mirela G Tulbure and Mark Broich, University of New South Wales, Sydney, NSW, Australia
Barmah-Millewa Forest (BMF), the largest River Red Gum forest in the world, located in south-eastern Australia is suffering from severe dieback, thus diminishing its ecological and economical value. Previous research showed that dieback is a good predictor of the forest health and stressed the need for BMF health mapping and change monitoring. In this respect, airborne laser scanning and hyperspectral imaging offer extensive spatial and spectral coverage of measurements and represent an ideal tool for forest health mapping at individual tree scale. The aim of this project is to quantify the health of individual, structurally complex floodplain eucalypt trees by integrating airborne hyperspectral imagery, full-waveform laser scans and field measurements. An aerial survey, conducted in May 2014, was designed to provide a representative sample of BMF tree health. The positioning of 17 flight lines aimed to capture the heterogeneity of the forest health and flood frequency. Preliminary analysis of the aerial remote sensing data with regards to chlorophyll concentrations, dieback levels and canopy densities allowed us to target our field campaign (conducted in June 2014). Field measurements included accurate position measurements, LAI, visual assessment, spectral measurement and mensuration of individual trees in 30 m2 plots. For detection of individual tree trunks from airborne laser scans we used a novel approach based on Euclidean distance clustering, taking advantage of the intensity and pulse width difference between woody and leaf tree compartments. The detected trunks were used to seed a minimum cut algorithm for tree crown delineation. In situ measurements confirmed the high structural diversity of the forest and allowed the calibration of the tree detection algorithm. An overall accuracy of the tree detection of 54% and 67% was achieved for trees with circumference over 40 cm and over 100 cm respectively. As a further step, 3D point clusters representing individual trees were used in conjunction with hyperspectral imagery to assess tree health in more detail. The detailed floodplain eucalypt tree health map, derived from aerial remote sensing data, will be further extrapolated to the whole forest using WorldView-2 imagery to produce a seamless, high resolution health map of the entire forest.