Multispecies deep learning using citizen science data produces more informative plant community models

Area of Science:

• Ecology
• Computational Biology
• Biodiversity Informatics

Background:

• Scientific progress is often limited by the ability to extract critical information from large datasets.
• Citizen science initiatives provide vast amounts of ecological data, but extracting fine-grained spatiotemporal information remains challenging.

Purpose of the Study:

• To develop and evaluate a novel approach using deep neural networks (DNNs) to map fine-grained spatiotemporal distributions of thousands of species.
• To compare the accuracy of multispecies DNNs against commonly-used methods for predicting species distributions and community composition.
• To explore the versatility of DNNs for investigating understudied ecological aspects like phenology and species dominance.

Main Methods:

• Utilized a large dataset of 6.7 million citizen science observations across Switzerland.
• Employed an ensemble of deep neural networks (DNNs) with varying cost functions to jointly model distributions of 2477 plant species and aggregates.
• Incorporated seasonal observation probabilities and cover-abundance data into the DNN models.

Main Results:

• Multispecies DNNs significantly outperformed traditional methods in predicting species distributions and, notably, community composition.
• The DNN approach successfully approximated flowering phenology by including seasonal observation variations.
• Reweighting predictions allowed for nationwide mapping of potentially canopy-dominant tree species.

Conclusions:

• Multispecies DNNs offer a powerful and versatile tool for refining our understanding of plant distributions and ecological dynamics.
• This methodology has the potential to advance ecological research, particularly for well-sampled taxa.
• Future projections using DNNs can assess the impact of environmental changes on species distributions, phenology, and dominance patterns.