Functional ecology has increasingly focused on describing ecological communities based on their traits (measurable features of individuals that affect their fitness and performance). Analyzing trait distributions within and among forests, could significantly improve understanding of community composition and ecosystem function. Historically, data on trait distributions in ecosystems is generated by (1) collecting samples of leaves from a relatively small number of trees, or (2) using remote sensing images to infer traits. The first method is limited by sparse and biased sampling, but produces information at the fundamental ecological unit: the individual. In contrast, remote sensing produces continuous information over large areas, typically with units of plots that contain dozens of trees of different species.
The National Ecological Observatory Network (NEON) provides data collected using both field collection and high-resolution remote sensing. Using these data we developed a method to scale up functional traits measured on 160 trees to the extent of NEON sites (millions of crowns). Our pipeline uses image segmentation algorithms to infer the size and position of individual tree crowns. We predict Leaf Mass Area (LMA), Nitrogen, Carbon, and Phosphorus content from the hyperspectral remote sensing using a naive multiple instance regression method based on ensemble of Partial Least Squares Generalized Linear models (pls-GLM). These models were fit at both the pixel (1m2) and the crown scale and used to predict the trait values of roughly 8 million trees at two NEON sites, Ordway Swisher Biological Station (OSBS) and Talladega (TALL).
Results/Conclusions
The trait models yielded Nitrogen, LMA and Phosphorus predictions with predictive R2 values ranging between 0.5 and 0.75 on held-out data, comparable with plot scale results from plot scale studies. An ensemble model resulted in higher predictability than single pls-GLM models. The crown delineation exhibited mixed performance, with overlap with ground truth crowns ranging from 25% to 78%, with oversegmentation being common. Despite these limitations, crown scale trait predictions outperformed pixel scale predictions, with improvements R2 ranging from 0.07 to 0.20 points.
Predicted traits at the two NEON sites showed expected patterns, such as a negative correlation between Nitrogen and LMA and bi-modal trait distributions.
Using crown segmentation algorithms to convert high resolution remote sensing images to information on individual trees appears to be beneficial both for improving model predictability. Converting remote sensing information to individual level predictions will also support combining remote sensing products to data collected from other ecological studies.