For most of its history, Earth observation has been an exercise in looking back. Satellites record what is there, analysts interpret it, and decisions follow from a picture of the recent past. That model has served the field well, but it is beginning to change. Increasingly, the questions put to Earth observation are forward-looking ones. Where is deforestation likely to advance next. Which fields are heading towards stress before it shows. How a landscape might respond to a flood, a fire or a shifting climate. The interesting frontier is no longer only what we can see, but what we can anticipate.
That shift runs into a constraint that gets discussed less than it should. Predictive and simulation-driven models are only as good as the data they learn from, and real satellite archives, for all their volume, are uneven. They are rich in ordinary, cloud-free, well-observed scenes and thin in exactly the rare or awkward conditions that matter most. When those conditions do appear, someone still has to find and label them by hand, slowly and imperfectly. The limiting factor, more often than not, is not imagery. It is usable training data.
This is the problem we have spent our time on at Another Earth, generating synthetic satellite data that behaves like real sensor data, so that models can learn from conditions the archive underrepresents. Much of that work has been about simulation and about detection tasks such as identifying forests and individual trees, the kind of capability that sits at the heart of very different industries. The common thread is that each one needs a model to recognise something real datasets barely contain, and generating those scenes ourselves means the labels are exact because we placed every object in them.
Brazil has become the clearest example of that for us. Our headquarters are in Austria, but we have recently opened operations in Brazil as well, drawn both by the scale of what is being built there and by how varied the demand for Earth observation turns out to be. The environmental questions are the ones the country is best known for, from deforestation and carbon monitoring across the Amazon to agriculture and water. Alongside them sits a growing set of industrial needs, from monitoring mining sites and pipelines to planning and inspecting the country's rapid rollout of solar installations. Few markets concentrate so many different uses in one place, and each one asks a model to learn something the archive barely holds.
Working in the tropics also means facing an old obstacle. At any moment roughly two-thirds of the planet sits under cloud, and over the tropics that share is higher still, so a great deal of the ground goes unseen precisely where change is fastest. This is the frontier we are focused on now, and the idea we are building around is counterintuitive: rather than only removing clouds, generate them. A synthetic scene can be rendered twice, once clouded and once clear, as a matched pair with perfect ground truth, giving cloud-detection and cloud-removal models a clean target to learn towards that real archives almost never provide.
If this direction holds, the implications for the industry are significant. The value of an Earth observation model would come to depend less on how much imagery already exists and more on how well the conditions it most needs to learn can be built. Real imagery stays essential, and synthetic data should always be measured against it. But as the field moves from monitoring towards prediction, the ability to build the missing data, deliberately and with exact ground truth, may turn out to be as important as the ability to capture it.