Google DeepMind Mapping, Modeling, and Understanding Nature with AI

Google DeepMind and Google Research have introduced a suite of AI-driven tools and research initiatives designed to monitor and protect the Earth’s biosphere. These advancements focus on three primary areas: predicting deforestation risk, scaling species range mapping, and automating bioacoustic monitoring.

Predicting Deforestation Risk with Vision Transformers

Google has developed a high-resolution model to predict the risk of deforestation, moving beyond simple tracking to proactive risk assessment.

Key technical details include:

  • Resolution and Scale: The model provides predictions of deforestation risk down to a 30-meter scale over large regions.
  • Architecture: The system is built around vision transformers and utilizes pure satellite inputs, eliminating the need for local input layers such as road maps.
  • Historical Context: This work builds on a model of forest loss drivers (including mining, logging, agriculture, and fire) developed with the World Resources Institute at a 1km² resolution for the years 2000․.

Google is releasing a benchmark dataset to support further research in forecasting forest loss.

Scaling Species Distribution Mapping via GNNs

To address the challenge of mapping the geographical ranges of millions of species, Google researchers are using a Graph Neural Network (GNN) approach to produce high-resolution species range maps at an unprecedented scale.

The GNN model integrates three primary data sources:

  1. Field Observations: Open databases of species sightings.
  2. Satellite Embeddings: Data from AlphaEarth Foundations.
  3. Species Traits: Biological information, such as body mass.

This methodology allows researchers to infer the underlying geographical distribution for multiple species simultaneously. In a pilot project with QCIF and EcoCommons, the model was used to map Australian mammals, including the Greater Glider. Twenty-three of these species maps have been released via the UN Biodiversity Lab and Earth Engine.

Bioacoustic Monitoring with Perch 2.0

Google has updated its animal vocalization classifier with the release of Perch 2.0. This model addresses the difficulty of manually reviewing the massive audio datasets generated by affordable bioacoustic monitors.

Capabilities and Applications:

  • Foundational Model: Perch 2.0 is designed as a foundational model, enabling field ecologists to quickly adapt it to identify new species and habitats globally.
  • State-of-the-Art Bird Identification: The model provides high-accuracy identification of bird vocalizations.
  • Conservation Impact: In collaboration with the University of Hawai`i, Perch 2.0 is being used to identify juvenile calls and guide protective measures for endangered honeycreepers.

Integrated Future of Biosphere Modeling

Google's long-term objective is to integrate these disparate models into a comprehensive system that combines multiple modalities—satellite data, images, bioacoustics, and documents—with models of human activity, such as land-use changes and agricultural practices. By combining environmental data with models of agricultural yields and flood prevention, Google aims to provide policymakers with a holistic understanding of biosphere threats to inform conservation and sustainability decisions.

Sources