Forecasting the future of forests with AI: From counting losses to predicting risk
Google Research ● Covered by 2 sources
Google built an AI model that predicts where forests will be cut down next, not just where they already were. It uses only satellite data, so it works anywhere and won't go stale like older road-and-population maps.
Based on reporting by Google Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Deforestation tracking has always been a rearview-mirror exercise. Satellites tell you a patch of Amazon or Congo Basin forest vanished last year, sometimes down to a square kilometer. That's useful for scorekeeping, but it does nothing for the community or government trying to stop the next hectare from disappearing. Google Research, working with Google DeepMind, just published a system called ForestCast that tries to flip the script: instead of counting losses after the fact, it forecasts where deforestation risk is highest before the chainsaws show up.
The scale of the problem explains the urgency. Tropical forests lost 6.7 million hectares last year alone, according to the paper, double the previous year's total and equivalent to wiping out 18 soccer fields every single minute. Land-use change of this kind accounts for roughly a tenth of global greenhouse gas emissions, and it's the single biggest driver of terrestrial biodiversity loss on the planet. Cattle ranching, palm oil, soy, illegal logging, mining, wildfire — the causes are tangled and region-specific, which is exactly why predicting them has been so hard.
Previous forecasting attempts leaned on a patchwork of inputs: road networks, population density, economic and policy data. That approach can work reasonably well in a specific place at a specific moment, but the maps go stale fast and have to be rebuilt region by region. Google's team instead trained a vision transformer using nothing but satellite imagery — Landsat and Sentinel-2 feeds, plus a derived layer they call
Read more about this at: Google Research