Automated assessment of the phenological state of herbarium specimens

Like many biological traits, phenology is inherently variable. It is therefore imperative to have large datasets from which patterns of phenological shifts can be discerned. Recent experiments in computer vision and machine learning have begun to automate the process of determining the reproductive state of a plant specimen. In this method, a computer (or more precisely a convolutional neural network) is ‘pre-trained’ to recognize flowers, fruits, or other phenological trait, and can then code the specimen accordingly, e.g., full flower. This pioneering work shows promise for upscaling our ability to gather phenological information from millions of specimens.