Nature underpins our local weather, our economies, and our very lives. And inside nature, forests stand as one of the crucial highly effective pillars — storing carbon, regulating rainfall, mitigating floods, and harboring nearly all of the planet’s terrestrial biodiversity.
But, regardless of their crucial significance, the world continues to lose forests at an alarming fee. Final yr alone, we misplaced the equal of 18 soccer fields of tropical forest each minute, totaling 6.7 million hectares — a file excessive and double the quantity misplaced the yr earlier than. At this time, habitat conversion is the best menace to biodiversity on land.
For years, satellite tv for pc information has been our important software for measuring this loss. Extra just lately, in collaboration with the World Sources Institute, we helped map the underlying drivers of that loss — from agriculture and logging to mining and hearth — for the years 2000–2024. These maps, that are at an unprecedented 1km2 decision, present a foundation for a variety of forest safety measures. Nonetheless these insights, crucial as they’re, solely look backward. Now, it is time to look forward.
We’re excited to announce the discharge of “ForestCast: Forecasting Deforestation Threat at Scale with Deep Studying”, together with the primary publicly accessible benchmark dataset devoted to coaching deep studying fashions to foretell deforestation threat. This shift from merely monitoring what’s already gone to forecasting what’s in danger sooner or later modifications the sport. Earlier approaches to threat have relied on assembling patchily-available enter maps, reminiscent of roads and inhabitants density, which might rapidly go old-fashioned. In contrast, we’ve developed an environment friendly method primarily based on pure satellite tv for pc information that may be utilized persistently, in any area, and may be readily up to date sooner or later when extra information turns into accessible. We discovered that this method might match or exceed the accuracy of earlier approaches. To make sure the neighborhood can reproduce and construct on our work, we’re releasing all the enter, coaching, and analysis information as a public benchmark dataset.