Skeletal trait measurements for thousands of bird species

Abstract

Large comparative datasets of avian functional traits have been used to address a wide range of questions in ecology and evolution. To date, this work has been constrained by the limited availability of skeletal trait datasets that include extensive inter- and intra-specific sampling. We use computer vision to identify and measure bones from photographs of museum skeletal specimens to assemble an extensive dataset of functionally important skeletal elements in birds. The dataset spans 2,057 species of birds (Aves: Passeriformes) and includes measurements of 12 skeletal elements from 14,419 individuals. In addition to the trait values directly measured from photographs, we leverage the multi-dimensional nature of our dataset and known phylogenetic relationships of the species to impute missing data under an evolutionary model. To facilitate use of the dataset, the taxonomy has been reconciled with an existing comprehensive avian phylogeny and an additional dataset of external functional traits for all birds.

Fig. 1

Dataset coverage. (A) The dataset includes species spanning all continents where passerines are resident. The ranges of all species included in the dataset are plotted, with colour indicating the number of species included in our dataset at each point in space. (B) The species in the dataset span Passeriformes, the most diverse taxonomic order within modern birds. Each bar associated with a tip on the phylogeny represents a species that is included in the dataset, with the height of the bar indicating the number of individuals of that species that were measured and included in the dataset, reflecting the high intra-specific sampling for some species and high variation in sampling among species. The bird skeleton highlights the bones that were measured in darker green.

Fig. 2

Variation in trait imputation accuracy across different levels of missing data. We masked an increasing percentage of the data generated by Skelevision (10–90% in intervals of 10%) and then imputed the missing values. We estimate the RMSE and P-Bias across ten randomized sets of data (comparing the imputed values to the Skelevision-generated values) for each percentage quantile and present the mean per trait. RMSE and P-Bias are generally low and stable across the range of missing data quantiles.