The nature of AI: solving the planet's data gap with Drew Purves
AI is being used to address critical information gaps in ecology. This summary covers how deep learning models like vision transformers and foundational models for sound are applied to map global forests, monitor deforestation, track species, and analyze bioacoustics to understand ecosystem health and animal communication.
A significant barrier to protecting and restoring natural ecosystems is a fundamental lack of information. While there is a global consensus on the importance of biodiversity, actions are often stalled by missing data about where species and habitats are, what condition they are in, and what the best course of action is. AI is poised to fill these critical information gaps.
The work in AI for nature can be seen in three key categories:
- AI for Data: Ingesting and processing data from diverse sources like field cameras, microphones, and scientific literature.
- AI for Information: Synthesizing this data with other sources, like satellite imagery, to create the derived information (e.g., maps) that decision-makers need.
- AI for Decision-Making: Deploying AI to help humans make sense of vast amounts of data and guide action.
Mapping the Biosphere
Much of the needed environmental information is geospatial, or simply, maps. Unlike the human world, the natural world is poorly mapped. There is no universally accepted, “gold-standard” global map for most habitats, including a basic distinction between natural and planted forests.
To address this, the Natural Forests of the World project has created a global map at 10-meter resolution that classifies each pixel with a probability of it being natural forest. This is powered by a multimodal, temporal, spatial vision transformer. This model architecture, adapted from language and vision applications, is uniquely suited to handle the complexities of satellite data, which includes:
- Massive scale: Enormous images with billions of pixels.
- Multimodality: Multiple data bands beyond RGB, such as infrared.
- Temporal dimension: Records from multiple satellites over many years.
- Missing data: Gaps in data due to cloud cover.
By providing a probability rather than a binary classification, the model allows downstream users to set their own confidence thresholds depending on their specific needs, democratizing access to high-quality remote sensing.
From Monitoring to Forecasting
By applying these mapping techniques to historical satellite data, it’s possible to understand not just where forests are, but how they have changed. A project with the World Resources Institute (WRI) maps the drivers of deforestation (e.g., logging, agricultural expansion) for every year since 2000 at a tenfold finer resolution than previous maps.
The next frontier is moving from retrospective analysis to real-time alerts and future forecasting. This is challenging because the future of human-driven processes like deforestation is not a simple extrapolation of the past. It can be influenced by sudden policy or market changes. The future may involve hybridizing machine learning with scientific simulation models and Large Language Models (LLMs) like Gemini, which could incorporate qualitative information from news and policy documents to create more robust, scenario-aware predictions.
The Sound of Ecology: Bioacoustics and Perch
For the vast majority of species not visible from space, other data sources are needed. Bioacoustics, the analysis of environmental sound, is an incredibly rich source of information. The Perch project is a foundational model for natural sound. Instead of being a pre-trained classifier for a fixed set of species, Perch creates a compact numerical representation—an “embedding”—of audio.
This embedding approach has several powerful applications:
- Custom Detectors: Ecologists can rapidly create their own detectors for rare species, or even distinguish between juveniles and adults, local “accents” in birdsong, or count individuals.
- Data Search: Researchers can find all similar-sounding clips in a massive audio dataset after finding just one or two examples.
- Ecosystem Health: The “audio ecology” of a place can serve as a proxy for its health. By comparing the soundscape embeddings of a restoration site to a healthy reference site (like a nature reserve), one can measure the progress of rewilding efforts over time. The stark difference in sound between a healthy and a degraded coral reef illustrates this potential.
Amazingly, the Perch model, initially trained on birdsong, works surprisingly well on underwater hydrophone data without retraining, suggesting there are universal, transferable features in natural communication sounds.
Decoding Communication and Changing Perspectives
The ultimate potential of this work extends beyond monitoring to understanding. Projects like DolphinGemma are using LLM-style architectures to begin decoding sophisticated animal communication. Successfully communicating with other species could fundamentally alter humanity’s relationship with the natural world, much like the first image of Earth from space or the sound of whale songs did. This shift in perspective may ultimately be AI’s most powerful contribution to conservation.