Determining the community composition of herbaceous species from images using convolutional neural networks

Matthias Körschens, Solveig Franziska Bucher, Paul Bodesheim, Josephine Ulrich, Joachim Denzler, Christine Römermann

Abstract: Highlights•Automatic prediction of the community composition reflects reference estimates well.•With annotations, the system can predict the composition better than human experts.•Without annotations, the system performs comparable to human experts.•Prediction works best for the most dominant species in the dataset.•Images from natural environments are better training data than herbarium specimens.