There is a category of AI problem that does not get enough attention in the mainstream conversation: the domains where the data needed to train useful models is structurally scarce. Not because the underlying phenomenon is rare, but because the act of observing and labelling it is too expensive, too slow, or physically impossible at the scale that good AI requires. Earth Observation is the clearest example of this problem at planetary scale. And Maya Pindeus has built a company specifically to solve it.
Another Earth generates synthetic satellite imagery: controllable, physically accurate, pre-labelled images that can be produced in any quantity, at any resolution, under any atmospheric conditions. The insight behind the company is direct. Satellites see the whole planet every day. The problem is not coverage — it is that the AI models needed to interpret what the satellites capture cannot be trained fast enough on real data alone. Rare events do not appear often enough. Labelling is expensive. Environmental conditions create gaps in coverage. Another Earth closes all three gaps at once by making the training data instead of finding it.
From Human-Machine Interaction to Planetary Intelligence
Before Another Earth, Pindeus co-founded Humanising Autonomy, a London-based startup working at the frontier of predictive AI and human-machine interactions. The company focused on making autonomous systems — particularly vehicles — safer by predicting human behaviour. She scaled it to over 50 people and raised more than $20 million in funding. It was serious, technically sophisticated work on a domain problem that millions of people eventually will care about.
The move to Another Earth was not a pivot away from hard problems. It was a move toward a harder one. Autonomous vehicles operate in bounded environments with relatively well-defined training domains. Earth Observation AI must generalise across every biome, every land use type, every atmospheric condition, every type of event that changes the surface of the planet. The training data requirements are orders of magnitude larger and more varied. And the consequences of getting it wrong — missed deforestation, undetected infrastructure damage, misread crop stress signals — play out at the scale of entire economies and ecosystems.
The Synthetic Data Approach
Another Earth's Synthetic Data Engine combines generative AI with procedural 3D simulation to produce satellite imagery that is not merely realistic but physically accurate — calibrated to the sensor characteristics of real satellite platforms, including their spectral bands, ground sampling distances, and atmospheric distortion profiles. The images are generated with ground truth annotations included by construction: every pixel is labelled, every object is classified, every scene variable is known. This is the property that makes synthetic data transformative for supervised learning at scale.
The applications span an unusually wide range of sectors. EUDR deforestation compliance monitoring requires detecting changes in forest cover across millions of hectares in near-real-time. Precision agriculture requires distinguishing between crop stress caused by drought, disease, or pest damage from orbit. Disaster response requires assessing infrastructure damage immediately after a flood or earthquake, before ground teams can reach the affected area. Each of these applications requires training data that real satellite archives cannot provide in sufficient volume or variety. Another Earth's platform generates it on demand.
The Founder Behind the Vision
Pindeus brings to Another Earth a background that spans design, engineering, and applied AI. She holds degrees from Imperial College London, the Royal College of Art, and completed the Masterclass Zaha Hadid at the University of Applied Arts Vienna — a combination that reflects the blend of technical rigour and design thinking that characterises Another Earth's approach to an engineering-heavy problem. She is a Forbes 30 Under 30 awardee, has contributed to the World Economic Forum, and has served on the UK Department for Transport's Expert Advisory Panel on autonomous systems.
The company closed a €3.5M seed round in early 2026, backed by Wake-Up Capital, Rockstart, Inovexus, and Stamco AG. It is at an early stage, but the problem it is solving — the training data bottleneck for Earth Observation AI — is one that grows more urgent with every percentage point of planetary change that outpaces the models designed to track it.
Implications
- For the climate and sustainability sector: AI-powered Earth Observation is not a future capability. It is deployable now, with synthetic data closing the gap between what satellites see and what models can be trained to understand. The bottleneck is no longer hardware or coverage — it is training data infrastructure. Another Earth removes that bottleneck.
- For AI practitioners in geospatial domains: The synthetic-to-real transfer problem in satellite imagery is more tractable than it appears, precisely because satellite sensors have well-characterised physics. A synthetically trained model calibrated to the right sensor profile can generalise to real imagery with less domain shift than equivalent approaches in other vision domains.
- For conference attendees: Pindeus brings to Human × AI a perspective that connects founding ambition, technical depth, and a genuine conviction that AI built around people and nature is not a constraint on performance — it is the condition for it.
Maya Pindeus joins Human × AI on May 19, 2026, in Vienna.