Ecosystem Mapping Approaches Based on Vegetation Structure Using NEON Prototype Airborne LiDAR and Field Data

Wednesday, 17 December 2014
Keith Krause1, William J Emery2, David Barnett1, Shelley Bougan Petroy1, Courtney L Meier1 and Carol Adele Wessman3, (1)National Ecological Observatory Network (NEON), Boulder, CO, United States, (2)Univ Colorado-CCAR, Boulder, CO, United States, (3)University of Colorado at Boulder, EBIO, Boulder, CO, United States
Remote sensing is a powerful tool for measuring the current state of vegetation and monitoring changes over time with repeated data collections. Airborne Light Detection and Ranging (LiDAR) data is especially well suited for mapping 3D vegetation structure.  In 2010, the National Ecological Observatory Network (NEON) contracted LiDAR and hyperspectral airborne data collections over the Ordway-Swisher Biological Station (OSBS). Ground truth campaigns were also conducted in 2010, 2011, and 2014 including structural measurements and generation of species lists for a set of ground validation plots. The vegetation communities at OSBS can be characterized by the Florida Natural Areas Inventory (FNAI) classification system, with a large area of the property belonging to the Sandhill community. For this study, classification algorithm training locations are hand selected for each FNAI community type using photo-interpretation. A series of LiDAR metrics are calculated on the discrete return point clouds and derived digital elevation (DEM) and canopy height models (CHM). A decision tree classification algorithm is run using R package “rpart”. A main goal of the project is to relate the LiDAR metrics used by the decision tree to direct canopy structural quantities. For instance, the canopy 75th minus the 50th percentile height in the LiDAR point clouds are related to the uniformity and light penetration in the upper canopy. A prototype of the decision tree achieved a classification accuracy of 89% on the training data itself, suggesting that some locations in different FNAI vegetation communities have similar structure and could not be distinguished in the LiDAR metrics used. An improved decision tree is currently under development which will include more training locations and more LiDAR metrics as input features. Results from this improved model will be presenting using the NEON ground truth locations as an independent and quantitative validation measure of the decision tree classification accuracy. Finally, a classification map is generated for entire site. Future work with the flux tower measurements, field vegetation observations, and airborne hyperspectral data from NEON at OSBS will facilitate analysis of how ecosystem structure is connected to ecosystem function.