Understanding species composition from NEON high resolution hyperspectral-LIDAR data across a heterogeneous landscape: Effects of land use, fire regime and topography

Abstract:

The 2014 NEON hyperspectral LIDAR data allows landscape scale analysis of how abiotic factors and management history affect ecosystem composition and function. At the Ordway Swisher Biological Station (OSBS) in Florida, the core terrestrial NEON site in the Southeast U.S., we are developing a framework applicable to all NEON sites for mapping and analyzing species composition. In this region, small changes in topography (elevation varies by only 20 m at OSBS) along with fire history are the dominant controls on tree species composition. To discriminate species, we use compare support vector machines (SVMs) with possibilistic classifiers (POCs), which may classify species as unknown and represent ambiguity among spectrally similar species better than common classifiers than (SVM). Species classification was most accurate (90%) using POC in the dominant upland longleaf pine forest type (where most trees belonged to just two species: Pinus palustris and Quercus laevis). It was lower (<60%) using SVM and in the hardwood hammocks (where > 10 hardwood species commonly co-occur, including multiple Quercus). For co-occurring hardwood species that were difficult to separate spectrally, we combined the NEON hyperspectral data with additional data sets (global and regional plant trait databases, state-level maps of ecosystem type, and U.S. Forest Service inventory data) in a possibilistic framework to increase the separability of species identity. We then generate a landscape-scale map of species composition at OSBS. Combining this species map with a LIDAR-derived topography, we show that species associations vary with topography. For example, some Quercus species tend to co-occur with each other in uplands, but not in mesic hammocks. We examine potential factors causing changes in community composition, including topography, water table depth, soil type, current fire management regime, and historical land use. By combining the NEON hyperspectral and LIDAR data with detailed records prescribed fire, we also examine how fire regime (frequency and year since last burn) affect fuel quantity and leaf chemistry.