loading page

Deep Images require Deep Learning: A Pixel-Based Convolutional Neural Network Classifier can Accurately Identify Tree Species Using Imaging Spectroscopy
  • +3
  • Geoffrey Fricker,
  • Janet Franklin,
  • Malcolm North,
  • Frank Davis,
  • Nicolas Synes,
  • Jeffrey Wolf
Geoffrey Fricker
California Polytechnic State University San Luis Obispo

Corresponding Author:africker@calpoly.edu

Author Profile
Janet Franklin
University of California, Riverside
Author Profile
Malcolm North
USDA Forest Service
Author Profile
Frank Davis
University of California
Author Profile
Nicolas Synes
One Degree North
Author Profile
Jeffrey Wolf
Columbia University of New York
Author Profile

Abstract

Our study uses field training data, airborne LiDAR (Light Detection and Ranging), imaging spectroscopy, and a Convolutional Neural Network (CNN) classifier to identify individual tree species in a mixed conifer forest in the Southern Sierra Nevada Mountains. The remote sensing data was collected on the National Ecological Observatory Network (NEON) Airborne Observation Platform in 2017. We trained the classifier using existing field plot data, and an independently collected validation dataset which identifies trees location of the 7 dominant species (Pine, Fir, Cedar and Oak), including condition and ‘live’ or ‘dead’ status. The LiDAR canopy height model was used to identify tree crowns and imaging spectroscopy data around these crowns were created as image ‘labels’. These species level ‘labels’ were used to train, test and validate a CNN tree species classifier. Our method achieved greater than 63-90% accuracy for all field validated stems and worked best for large diameter trees. On an independent tree stem dataset, we performed a species-level logistic regression to study which cases the classifier works most and least effectively. Spatially, in the southern areas scattered Black Oak were present and tree species often were confused with shrub species or covered by adjacent conifer species. In the north where the upper elevation forest is dominated by red and white fir the classifier achieved greater than 96% accuracy for larger canopy trees, with accuracy degrading to about 59-70% when smaller trees are included in the model. There is also genus level mis-classification, particularly between Red and White Fir species. There was high tree mortality in this forest and the classifier was effective in detecting large tree mortality, which also varies as a function of species and size of trees. This work leverages newly developed ‘deep learning’ tools which have yet to be extensively applied to the remote detection of large trees or in plant biogeography generally. This research is a proof-of-concept for forest community ecologists who want accurate ‘tree species maps’ to study how plants are distributed across space. Generally, this method will be of interest to biogeographers or remote sensing scientists looking to apply novel classification methods to problems beyond remote tree species identification.