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 the vast majority of the planet’s terrestrial biodiversity.
But, regardless of their vital significance, the world continues to lose forests at an alarming price. Final yr alone, we misplaced the equal of 18 soccer fields of tropical forest each minute, totaling 6.7 million hectares — a document excessive and double the quantity misplaced the yr earlier than. Immediately, habitat conversion is the best menace to biodiversity on land.
For years, satellite tv for pc information has been our important device for measuring this loss. Extra not too long ago, 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, vital 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 out there benchmark dataset devoted to coaching deep studying fashions to foretell deforestation danger. This shift from merely monitoring what’s already gone to forecasting what’s in danger sooner or later adjustments the sport. Earlier approaches to danger have trusted assembling patchily-available enter maps, comparable to roads and inhabitants density, which might shortly go outdated. Against this, now we have developed an environment friendly strategy primarily based on pure satellite tv for pc information that may be utilized persistently, in any area, and will be readily up to date sooner or later when extra information turns into out there. We discovered that this strategy might match or exceed the accuracy of earlier approaches. To make sure the group can reproduce and construct on our work, we’re releasing the entire enter, coaching, and analysis information as a public benchmark dataset.
