Google DeepMind unveiled three new AI projects designed to help scientists map, model, and protect Earth's ecosystems. The initiatives tackle deforestation prediction, species distribution mapping, and bioacoustic monitoring—areas where machine learning can process data at scales impossible for human researchers.

The first project, developed with the World Resources Institute, predicts deforestation risk at 30-meter resolution using vision transformers trained on satellite imagery. The benchmark dataset covers forest loss drivers like agriculture, logging, and mining across 1km² areas from 2000-2024, eliminating the need for local input layers like road networks.

Researchers are also using graph neural networks to map species ranges at unprecedented scale. The model combines field observations, satellite embeddings from AlphaEarth Foundations, and species traits to infer geographical distributions. A pilot with QCIF and EcoCommons mapped Australian mammals including the Greater Glider, with 23 species maps released through the UN Biodiversity Lab. Meanwhile, Perch 2.0—the latest version of Google's animal vocalization classifier—now functions as a foundational model, allowing ecologists to quickly adapt it for identifying new species. The University of Hawaiʻi is using Perch to protect endangered honeycreepers by identifying juvenile calls and monitoring population health.

The goal extends beyond individual models: Google aims to integrate satellite data, bioacoustics, and other modalities alongside human activity models to give policymakers a comprehensive view of ecosystem threats.