Automated extraction of leaf mass per area from digitized herbarium specimens

The digitization of vast herbarium collections has made millions of plant specimen images freely available online, which can now be used to generate phenotypic datasets of unprecedented scope. Here, we assess the potential of computer vision tools to automate the extraction of predicted leaf mass per area (LMA_pred) from digitized herbarium specimens. We use an automated pipeline to extract leaf area and petiole width from 22 680 leaves, representing a phylogenetic informed sample of 1580 species of woody angiosperms. LMA_pred is estimated using a proxy equation that models the scaling relationship between petiole width and leaf mass. We assess potential sources of error in LMA_pred estimates and evaluate whether documented LMA-climate patterns are recovered using this dataset and phylogenetic comparative methods. Our LMA_pred dataset responds mainly to temperature and solar radiation and presents a positive correlation with latitude. The proxy equation, not the automated pipeline, is responsible for most of the error in LMA_pred estimates. Our pipeline underscores the power of combining herbarium digitization with new techniques for automated trait scoring. The increased size of datasets generated using this tool allows investigation of potential LMA-climate relationships with a geographically balanced sample while also utilizing comprehensive phylogenetic information.