The University of California, Berkeley has made significant strides in mapping plant distributions across California using advanced deep learning methods. This breakthrough comes via a collaboration that taps into a popular citizen science app called iNaturalist. Originally conceived at UC Berkeley, iNaturalist allows users to upload photos and location information of various life forms they encounter, with the community then assisting in identifying these sightings. With over 200 million observations uploaded by more than 8 million global users, this app has become a pivotal tool for ecological research.
The researchers at UC Berkeley harnessed this wealth of data using a deep learning model known as a convolutional neural network, a type of artificial intelligence that excels in handling image data. The model, termed "Deepbiosphere," analysed data from iNaturalist alongside high-resolution satellite and aerial images, allowing for precise predictions of the ranges of 2,221 plant species throughout California, down to areas as small as a few square metres.
Traditionally, botanists have had to painstakingly review specific small areas to build species distribution maps, a method impractical over larger scales. Deepbiosphere offers a scalable alternative, effectively automating the mapping process by leveraging freely available data from both iNaturalist and remote sensing imagery from sources like satellites or airplanes.
This innovative approach was detailed in a study published on 5th September in the Proceedings of the National Academy of Sciences. The study was led by Moisés “Moi” Expósito-Alonso, an assistant professor of integrative biology at UC Berkeley, and first author Lauren Gillespie, currently a doctoral student at Stanford University. Gillespie, who is also working under a Fulbright grant to apply similar techniques in Brazil, highlighted the practical applications of the work. In Brazil, experiencing severe droughts and fire seasons, such methods could elucidate impacts on individual species, using deep learning to track environmental changes in real-time.
The potential of Deepbiosphere extends globally. "Our dream," Expósito-Alonso shared, "is to integrate such data into widely used platforms like Google Maps, allowing for the protection of diverse species worldwide."
Deepbiosphere’s efficiency was put to the test against existing models, including Maxent, a tool developed by the American Museum of Natural History, which uses climate and species occurrence data. Deepbiosphere demonstrated superior accuracy, predicting plant presence with 89% accuracy compared to Maxent’s results. Further, the model was successfully employed in mapping complex ecosystems, such as California’s Redwood National and State Parks, illustrating its capability in distinguishing between various growth stages of redwood forests.
Furthermore, apart from being free and accessible, remote sensing data offers high granularity and frequent updates, enabling detailed monitoring of landscape changes, often elusive but now possible to detect with Deepbiosphere. This capability could revolutionise how conservationists approach areas experiencing environmental stress or biodiversity shifts.
The model’s proficiency was again highlighted through trials in Yosemite National Park, where it accurately modelled the impacts of the severe 2013 Rim Fire, offering insights into burn severity and plant community transformations at a resolution of 35 metres.
This work, funded by the Carnegie Institution for Science, began at Stanford before Expósito-Alonso’s move to UC Berkeley. Megan Ruffley from Stanford's Carnegie Institution was also a contributing co-author. Through this project, the burgeoning intersection of citizen science, AI, and conservation research demonstrates immense possibilities for understanding and preserving global biodiversity in a changing climate.
Source: Noah Wire Services