Google's New AI Models Predict Deforestation and Map Species

Google has released a deforestation-risk prediction benchmark, a new Graph Neural Net approach for mapping the ranges of Earth's species, and Perch 2.0, an updated animal vocalization classifier — a set of tools the company says can help governments, companies and conservation groups turn field data into actionable protection plans.

The announcements come from Google DeepMind and Google Research, and arrive as demand for land and resources intensifies pressure on the ecosystems that produce the air, water and food humans depend on. Google frames AI as a way to make conservation cheaper and faster: collect field data more easily, integrate it into insights, and monitor whether protection plans actually work.

Predicting deforestation at 30-meter resolution

Forests store carbon, regulate rainfall, mitigate floods and harbor the majority of the planet's terrestrial biodiversity, yet they continue to be lost at what Google calls an alarming rate. Satellite-based remote sensing has made it possible to track deforestation from space for more than 20 years.

Together with the World Resources Institute, Google went one level deeper. The team built a model of the drivers of forest loss — from agriculture and logging to mining and fire — at an unprecedented 1km² resolution, covering the years 2000-2024.

The benchmark dataset released today uses pure satellite inputs, avoiding the need for local data layers such as road maps, and relies on an efficient architecture built around vision transformers. According to Google, this approach enables accurate, high-resolution predictions of deforestation risk down to a scale of 30 meters, across large regions. For policymakers and conservation groups, that granularity could shift deforestation work from documenting loss after the fact to anticipating where it will happen next.

Mapping where species live

To conserve threatened species, scientists first have to know where they are. With more than 2 million known species and millions more to be discovered and named, Google calls that a monumental task.

The company's answer is a Graph Neural Net (GNN) model that combines open databases of field observations with satellite embeddings from AlphaEarth Foundations and species trait information such as body mass. The model infers likely geographical distributions for many species at once — more species, over more of the world, at higher resolution than before, Google says — and scientists can then refine those inferred maps with local data and expertise.

In a pilot with researchers at QCIF and EcoCommons, Google used the model to map Australian mammals including the Greater Glider, a nocturnal, fluffy-tailed marsupial that lives in old-growth eucalyptus forests. Twenty-three of these species maps are being released today via the UN Biodiversity Lab and Earth Engine, making them directly available to the institutions that advise biodiversity policy.

Listening through bioacoustics

Field monitoring is the foundation of all ecosystem modeling, and it is notoriously difficult and costly. Birds, amphibians, insects and other species communicate by sound, which makes bioacoustics a strong modality for identifying resident species and gauging ecosystem health. Affordable bioacoustic monitors are readily available — but the devices generate vast audio datasets full of unknown and overlapping sounds, too large to review manually and hard to analyze automatically.

Perch 2.0, Google's recently released update to its animal vocalization classifier, is the company's attempt to solve that bottleneck. Google says the model is state of the art for bird identification and is also available as a foundational model, letting field ecologists quickly adapt it to identify new species and habitats anywhere on Earth — a shift from fixed classifiers tuned to well-studied regions toward a general-purpose tool.

The model is already deployed in conservation work. At the University of Hawai`i, Perch guides protective measures for endangered honeycreepers and identifies juvenile calls to help researchers understand population health.

Why it matters

The stated goal is to make it easier for decisionmakers at all levels to act. But Google acknowledges that better data only leads to better decisions if the data is thorough and truly captures what is happening in an ecosystem at all levels.

That limitation explains the company's next step: integrating these and other models together, combining satellite data, images, bioacoustics and documents, and joining them with models of human activity such as land-use changes and agricultural practices, as well as models of agricultural yields and flood prevention. By giving policymakers a comprehensive view of threats to the biosphere, Google argues, AI can help protect future generations of plants, animals and people. As the company puts it: "If we can model the environment, perhaps we can help it thrive."