Using machine learning to predict plant trait distributions from citizen science photos

Combining trait data with observational photographs, study creates neural network model for automated predictions of global distributions of plant traits

GBIF-mediated data resources used : 10+ M species occurrences
Calceolaria hyssopifolia Kunth observed in Cuenca, Ecuador by bb_593 (CC BY-NC 4.0)

How ecosystems respond to global change can best be assessed by examining the functional traits available among species within the ecosystem. Rapid ecosystem monitoring could be facilitated by effective community trait measurements aided by machine learning techniques.

In this paper, authors designed a Convolutional Neural Network (CNN) aimed at predicting plant traits from photographs, trained using the TRY database of known species traits, e.g. leaf area, seed mass and stem density, combined with GBIF-mediated iNaturalist observations with photographs.

Adding data on trait plasticity and local climate at the location of an observation, the authors created an ensemble model showing high replicability and low error rates, generalizing well across growth forms, taxa and biomes.

To demonstrate the potential of the model, the authors used a separate set of iNaturalist observations to derive a global map of plant trait distributions. The produced map expressed a unimodal latitudinal distribution peaking around the equator for traits such as leaf area, height and seed mass, while leaf nitrogen concentration also showed peaks in northern temperate and polar zones.

In North America, a longitudinal gradient of high (in the east) to low (in the west) leaf area was apparent, while the opposite trend was observed for nitrogen concentration and stem density. These predictions combined correlated significantly with existing global maps of plant traits.

Schiller C, Schmidtlein S, Boonman C, Moreno-Martínez A and Kattenborn T (2021) Deep learning and citizen science enable automated plant trait predictions from photographs. Scientific Reports. Springer Science and Business Media LLC 11(1). Available at: