Google $GOOGL Earth looks detailed enough to count cars in a parking lot. That makes it natural to wonder why companies building AI for robots, autonomous vehicles, and augmented reality can't just use satellite imagery to train their models.
The answer comes down to a technical concept called ground sample distance, which measures the physical distance each pixel in an image represents. Spexi, one company working on the problem, delivers drone imagery with a resolution of 2.8 centimeters per pixel. The best commercial satellites offer about 30 centimeters of resolution. That factor-of-ten gap is the difference between seeing a sidewalk crack and seeing a gray smear.
What satellites can and can't resolve
The resolution ceiling for commercial satellite imagery has been shaped as much by regulation as by optics. NOAA, the U.S. regulatory agency overseeing commercial remote sensing, long limited the resolution of commercially available satellite images — the federal limit was relaxed from 50 centimeters to 25 centimeters in 2014.
The highest-resolution commercial satellites in orbit reflect that history. Maxar's WorldView-3 collects 31-centimeter panchromatic imagery. Planet Labs' SkySat constellation captures imagery at up to 50 centimeters per pixel. Planet's daily-monitoring PlanetScope constellation, which images all of Earth's land surface every day, operates at a resolution of about three meters per pixel.
In March 2025, Albedo Space launched Clarity-1, a satellite designed for Very Low Earth Orbit. The company holds a NOAA license to acquire commercial imagery with resolutions of 10 cm panchromatic and 40 cm multispectral. Clarity-1 proved the viability of sustained VLEO operations but did not deliver commercial imagery at scale, serving as a pathfinder rather than a production system. Even so, the resolution target, if achieved at scale by future satellites in the constellation, would significantly narrow the gap with aerial data. But even 10 cm per pixel remains about four times coarser than what standardized drone flights produce.
Why resolution is only half the problem
For AI training, consistency matters as much as clarity. An AI model trained on imagery captured by different sensors, at different altitudes, under different lighting conditions, and with different geometric distortions will learn those inconsistencies as features of the world rather than artifacts of data collection. Each variation introduces a domain shift, where the statistical properties of the training data differ from those of the inference data, creating reliability challenges.
A model's sensitivity to ground sample distance represents one of the most significant technical barriers in geospatial AI, according to research published by Kili Technology. A model trained on 30 cm imagery and deployed on 3 cm imagery faces a tenfold shift in what each pixel represents. Objects that occupied a single pixel in training data now span a hundred pixels. The model has no framework for interpreting that difference.
Given the specific nature of geospatial data, standardizing and adjusting data is essential for producing accurate GeoAI models — data must be reliable, normalized, and uniform in format and structure, according to a comprehensive GeoAI review published on arXiv. Satellite imagery gathered from different providers, at different times, with different sensors violates those requirements. Traditional aerial surveys present similar problems: the data comes from various contractors using different equipment under different conditions.
This is the context for the broader shift in AI training data. The data layer that geospatial AI companies need, combining high resolution with standardized capture methods, did not exist. Companies like Spexi are trying to build it by deploying a network of more than 10,000 drone pilots flying missions at an altitude of 80 meters with uniform equipment and automated flight plans. The company's orthomosaics capture 2.8 cm ground sample distance imagery that is fully calibrated and standardized, and ready for ingestion into modeling pipelines, according to a February 2026 announcement from its distribution partner, SkyWatch.
What these models are actually for
The demand for this level of detail is coming from a specific set of applications. Large geospatial models will enable computers not only to perceive and understand physical spaces but also to interact with them, forming a critical component of AR glasses and fields beyond, including robotics, content creation, and autonomous systems, according to Niantic.
Niantic Spatial, the geospatial AI company spun off from the Pokémon Go creator in May 2025, has built its Visual Positioning System on more than 50 million neural networks with over 150 trillion parameters, enabling operation across over a million locations. The company's Large Geospatial Model is built on a proprietary database of over 30 billion posed images. That system requires centimeter-level positioning accuracy, the kind of detail that satellite imagery cannot deliver.
The use cases make the resolution requirements concrete. Niantic Spatial can pinpoint a user's exact location on a map to within a few centimeters based on landmarks and buildings within a camera's field of view, according to PetaPixel. A delivery robot navigating a sidewalk needs to distinguish between a curb, a planter, and a bench. An AR application overlaying digital objects on a real-world street scene needs sub-centimeter alignment. A 30 cm satellite pixel would render all three objects as the same undifferentiated blob.
The trade-off that matters
Satellites retain clear advantages. When it comes to coverage area, satellite imagery far surpasses drone imagery, making it the preferred choice for large-scale geographic analysis. Satellites can access all geographical locations regardless of extreme terrain or weather conditions. For tracking deforestation, monitoring crop health across regions, or observing urban sprawl at a continental scale, satellite data remains the right tool.
But the AI applications driving the next wave of demand don't need to see everything on Earth at low resolution. They need to see specific environments at a level of detail that matches the tasks machines will perform in them. That is a different product, and no amount of satellite coverage can substitute for it. The question for the geospatial AI market is whether standardized, on-demand drone capture can scale fast enough to meet the demand.