Scientists use machine learning to predict forest tree species diversity

A collaborative research team led by Ben Weinstein of the University of Florida in Oregon used machine learning to create a detailed map of more than 100 million trees in 24 locations across the United States, and published their findings in an open-access journal on July 16. PLOS BiologyThese maps provide information on the species and condition of individual trees and are extremely useful for conservation efforts and other ecological projects.

Ecologists have been collecting data on tree species for years to better understand forests' unique ecosystems. Historically, this was done by surveying small plots of land and extrapolating the results, but this cannot account for variation across the forest. Other methods can cover larger areas but often make it difficult to classify individual trees.

To create the large-scale, highly detailed forest map, the researchers used aeroplane-captured tree canopy images and other sensor data to train a type of machine learning algorithm called a deep neural network. These training data covered 40,000 individual trees and, like all data used in the study, were provided by the National Ecological Observatory Network.

The deep neural network was able to classify the most common tree types with 75-85% accuracy. Additionally, the algorithm can provide other important analytics, such as reporting which trees are alive and which are dead.

The researchers found that the deep neural network showed the highest accuracy in areas with more open space within the tree canopy, and performed best at classifying coniferous tree species such as pines, cedars, and sequoias.

The network also performed best in areas with low species diversity. Understanding the strengths of the algorithms will help in applying these methods to different forest ecosystems.

The researchers also uploaded their model predictions to Google Earth Engine, making their findings useful for other ecological studies. “The diversity of overlapping datasets will enhance our understanding of forest ecology and ecosystem functioning,” the researchers say.

The authors add: “Our goal is to provide researchers with the first wide-area maps of tree species diversity in ecosystems across the United States. These canopy tree maps can be updated with new data collected at each site, and by collaborating with researchers across NEON sites, we can build better predictions over time.”