You are here

Harvard Forest >

Harvard Forest Symposium Abstract 2019

  • Title: Using Airborne Hyperspectral and LiDAR Remote Sensing to Map Tree Species Distribution at Harvard Forest, Massachusetts, USA
  • Primary Author: Jack Hastings (University of New Hampshire - Main Campus)
  • Additional Authors: David Basler (Harvard University); Scott Ollinger (University of New Hampshire - Main Campus); Andrew Ouimette (University of New Hampshire - Main Campus); Michael Palace (University of New Hampshire - Main Campus); Franklin Sullivan (University of New Hampshire - Main Campus)
  • Abstract:

    The ability to classify species at the individual tree crown level using high resolution remote sensing is an important step in scaling species – specific information to whole ecosystems. In closed canopy temperate forests, efforts to achieve this goal are hampered by the challenges of 1) distinguishing multiple tree species that have similar spectral properties and 2) automating tree crown delineation.
    To address these challenges, we combined high resolution hyperspectral and LiDAR remote sensing data collected by the National Ecological Observatory Network (NEON) to map and classify individual trees across the Harvard Forest, in Petersham, Massachusetts. Using forest inventory data from the ForestGEO MegaPlot, and multi-temporal UAV imagery, we manually delineated over 1300 crowns from 7 dominant tree species. Characteristic combinations of functional (hyperspectral-derived) indices and structural (LiDAR-derived) metrics were extracted from each crown and used to train and validate machine learning classifiers for species identification. We evaluated five different automated LiDAR delineation methods for their ability to successfully identify individual tree crowns. Automated methods were compared against manually delineated crowns (figure 1) in 15 plots that varied in species composition and structure.
    Our results show that individual species can be identified with high accuracy (>90% overall accuracy) using spectral reflectance data. Further, combining remotely sensed spectral and LiDAR metrics improves the classification of species due to distinct species-specific structural characteristics (figure 2). Despite our success in classifying the identity of pre-delineated tree crowns, automated detection of individual tree crowns using standard LiDAR-based approaches remains difficult and unreliable (<50% overall accuracy). The success of automated delineation methods using structural (LiDAR) data alone varied across stands with different species composition and structure. They performed best in hemlock and pine dominated stands; stands with high canopy structure and distinct, uniform crown shape. Delineation of hardwood canopies often resulted in over- or under- segmentation because of crown overlap and plasticity, and generally lower canopy structure. Integrating spectral and phenological characteristics with LiDAR would likely improve automated ITC delineation in dense, mixed-hardwood stands.

  • Research Category: Biodiversity Studies