Abstract
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 are generated by (1) collecting a few leaves from a small number of trees, or (2) using remote sensing images to infer traits. While the first method suffers from limited, and potentially biased, sampling, it produces information at the level of the fundamental ecological unit: the individual. On the contrary, remote sensing produces information in a continuum space whose unit is usually the plot, containing dozens of trees of different species.
The National Ecological Observatory Network (NEON) provides data collected both from the field and 1m2 resolution remote sensing. Using these data we developed a method to scale up functional traits measured on 160 trees to the full spatial extent of NEON sites (millions of crowns). The first step of our pipeline involved the use of image segmentation methods to infer crown geometry. Then, we built models to infer Leaf Mass Area (LMA), nitrogen, carbon, and phosphorus content. We developed a naive multiple instance regression method based on an ensemble of Partial Least Squares Generalized Linear regressions (pls-GLR). Results were applied both at the pixel scale (1 m2) and aggregated to the crown scale for comparison. Finally, we applied the models to two NEON sites, Ordway Swisher Biological Station (OSBS) and Talladega (TALL) to estimate the location and leaf traits of approximately 5 million trees.
Nitrogen, LMA and phosphorus models had R2 values ranging from 0.50 to 0.75 on held out test data, comparable with what has been observed in studies conducted at the plot scale. The ensemble resulted in better predictability than single pls-GLM models. Models for carbon performed poorly compared to other traits (R2 of 0.22). Crown delineation produced the highest uncertainty in the pipeline, generally due to over-segmentation of crowns. The intersection over union between segmented and ground-truth crowns ranged between 0.24% and 0.78%. Despite this uncertainty, crown scale predictions performed significantly better than pixel predictions, resulting in improvement of R2 on test data of between 0.07 to 0.20 points.
We used the resulting models to produce individual trait data for ~2.5 million individual crowns for both OSBS and TALL, covering a total extent of ~360 km2. In these “derived-data” LMA, %N, and %P were highly correlated as expected. Chemical traits were highly correlated to some physical traits like LAI, and to elevation. Structural traits like crown area and height were correlated to chemical traits only in OSBS.