An image is worth a thousand species: mapping plant biodiversity with citizen science, remote sensing, and deep learning

Anthropogenic habitat destruction and climate change are altering the composition of plant communities worldwide1,2. However, traditional species distribution models cannot detect rapid, local plant species changes due to their low spatial and temporal resolution3,4, and remote sensing models can only identify changes in coarse vegetation categories5,6. Here we combine open-access remote sensing imagery, citizen science observations, and deep learning to create a multi-species prediction model at high spatial and temporal resolution. We train a novel deep convolutional neural network using [~]half a million observations within California to simultaneously predict the presence of over 2,000 plant species at meter-level resolution. This model--deepbiosphere--accurately performs many key biodiversity monitoring tasks, from fine-mapping geographic distributions of individual species and communities, to detecting rapid plant community changes in space and time. Deepbiosphere shifts the paradigm for species distribution modeling, providing a roadmap for inexpensive, automatic, and scalable detection of anthropogenic impacts on species worldwide.