Is Your GeoAI Solving for Earth or for Flat Maps?
Levente Juhász is an Assistant Professor of Geospatial Analytics at the University of Florida and director of the Geospatial Analytics, Technology and Open Research (GATOR) Lab. At the H3 Conference 2026 in New York City, Levente argued that GeoAI models built on flat, pixel-based maps carry projection distortion into their answers, and that discrete global grid systems (DGGS) such as H3 are a better foundation, as long as you understand their trade-offs.
Flat maps distort, even inside one UTM zone
Levente is a geographer by training. Noticing that foundation models such as AlphaEarth still rely on flat, pixel-based maps, Levente measured the distortion of the UTM projection that AlphaEarth uses across Florida.
Even UTM, which is precise locally by design, shows it:
- A 10 m by 10 m pixel near St. Augustine, at the center of the zone, covers more than 100 square meters of the Earth's surface.
- The same pixel near Tallahassee, close to the zone edge, covers less than 100 square meters.
- The difference is a 0.23% area discrepancy. That sounds like nothing, but for 100 m by 100 m pixels it becomes 23 square meters per pixel, about the size of a large garden shed.
No one, Levente said, has looked into whether this makes global analyses underestimate or overestimate things. Flat maps also break at their edges: places in Fiji that sit next to each other split apart when the map is treated as a continuous coordinate space. In Levente's words, we are forcing geo models to solve problems that don't exist natively on the sphere.
Flat maps made sense in the era of the Mercator projection, paper maps and early computers. Today, Levente argued, they are largely technical legacy, and we have the compute and the math to move to data structures that might give more accurate, less biased answers.
How discrete global grids are built
A DGGS starts from a base solid, usually a platonic solid such as an icosahedron or a cube, projects it onto a sphere or ellipsoid, and reconstructs the cells. H3 uses hexagons, Google's S2 uses quadrilaterals, and pentagonal systems exist too. All of them offer topological consistency across the globe: no edges, cells that work at the poles, and spatially and temporally stable cell identifiers.
Benchmarking H3 against other grids
Levente built a simple Python benchmarking utility that abstracts the leading DGGS implementations so experiments can be plugged in. The talk abstract lists S2, A5, ISEA and HEALPix variants among the grids compared with H3. Two of the experiments stood out:
- Area and angular distortion. H3 cells are not constant in area, so H3 alone does not solve the UTM problem above, while some other DGGS are fully equal area by design. Hexagonal grids like H3 are smooth on angular distortion because of the hexagon's regularity.
- Spatial join speed. Joining on cell IDs, implemented properly, is far faster than the usual coordinate geometry approach with
ST_Intersects. Levente advocates moving science from desktop GIS approaches to faster, cloud-native ones.
H3 is the most widely adopted DGGS because of its tooling: it works out of the box in DuckDB through a community extension, has an optimized core, and is fast and easy to learn. The equal-area "scientific" grids lack that tooling, which holds back their adoption.
H3's area variation matters for analysis
The caution for H3 users: it is not an equal-area grid, and cell sizes vary by more than 10% by design. That doesn't matter for visualization and many other uses, but it matters for analytical work.
To show the effect, Levente intersected Foursquare Open Source Places, more than 100 million POIs worldwide, with H3 cells, ranked the cells by POI count, then ranked them again taking each cell's area into account. More than 80% of the cells moved by over 1% in the ranking just because of the area mismatch.
The conclusion: H3 is currently the best-supported DGGS and works for a wide range of use cases, but academia should pay more attention to these grids, especially equal-area ones.