Inferensys

Glossary

H3 Hexagonal Grid

A discrete global grid system developed by Uber that partitions the Earth's surface into hierarchical hexagonal cells, providing a standardized, distortion-minimizing spatial indexing system for aggregating and querying Radio Environment Map data.
Developer building agentic RAG system, retrieval pipeline diagram on laptop, technical workspace with notes.
DISCRETE GLOBAL GRID SYSTEM

What is H3 Hexagonal Grid?

A hierarchical geospatial indexing system that partitions the Earth into hexagonal cells to standardize the aggregation and querying of radio environment map data.

An H3 Hexagonal Grid is a discrete global grid system (DGGS) developed by Uber that partitions the Earth's surface into a hierarchical set of hexagonal cells. Each cell is assigned a unique 64-bit integer index, providing a standardized, distortion-minimizing spatial reference for aggregating geospatial data, such as spectrum occupancy measurements, across varying resolutions without the angular distortion inherent in square grids.

In Radio Environment Mapping, H3 enables efficient spatial indexing and querying of heterogeneous RF sensor data. Its hexagonal topology ensures uniform adjacency—every cell has six equidistant neighbors—eliminating the diagonal connectivity ambiguities of rectangular grids. This property is critical for accurate spatial interpolation and propagation modeling, allowing systems to seamlessly aggregate signal power, interference levels, and spectrum opportunity data at multiple hierarchical resolutions.

Spatial Indexing Architecture

Key Features of H3 for Spectrum Data

H3 provides a hierarchical, hexagonal discrete global grid system that partitions the Earth into cells of uniform area and shape, enabling distortion-minimizing aggregation and querying of radio environment map data.

01

Hierarchical Hexagonal Partitioning

H3 partitions the globe into 12 base pentagons and a recursive subdivision of hexagonal cells across 16 resolution levels. Each cell is uniquely identified by a 64-bit integer index, enabling deterministic spatial referencing. The hexagonal geometry ensures uniform adjacency—every cell has six equidistant neighbors—eliminating the diagonal connectivity ambiguities inherent in square grids. Resolution ranges from 1,107 km² at level 0 to 0.9 m² at level 15, allowing spectrum maps to scale from continental overviews to micro-cell propagation analysis.

16
Resolution Levels
0.9 m²
Finest Cell Area
02

Distortion-Minimizing Area and Shape

Unlike traditional geohash or S2 square grids that suffer from significant area distortion at high latitudes, H3 hexagons maintain minimal variation in cell area across the globe. The maximum area ratio between the largest and smallest cell at any given resolution is kept below 1.4x, compared to orders of magnitude for equal-angle grids. This property is critical for spectrum occupancy heatmaps, where signal power density calculations require consistent spatial binning to avoid statistical bias in aggregated RF measurements.

< 1.4x
Max Area Variation
04

Grid-Based Kriging and Spatial Interpolation

H3 cells serve as the canonical spatial support for geostatistical interpolation in REM construction. Kriging and Gaussian Process Regression models compute variogram parameters over H3 cell centroids, treating each hexagon as a discrete measurement support. The uniform cell area ensures that sensor density normalization—dividing aggregated signal counts by cell area—is consistent across latitudes. This enables accurate spectrum cartography where predicted power spectral density values are assigned to H3 indexes for efficient spatial querying.

05

Polygon-to-Cell Conversion and Coverage Mapping

H3 provides polyfill and polygonToCells functions that convert arbitrary geofences—such as exclusion zones, incumbent protection contours, or building footprints—into sets of H3 indexes at a specified resolution. This enables:

  • Spectrum Access System (SAS) enforcement by pre-computing prohibited transmission cells
  • Propagation coverage maps by intersecting ray-tracing output polygons with H3 grids
  • 3D city model integration by mapping building geometries to H3 cells for diffraction loss lookup tables
06

Neighbor Traversal and Graph Neural Networks

The gridDisk and gridRing functions return the set of H3 indexes at a specified k-ring distance from a central cell. This deterministic adjacency structure makes H3 an ideal graph topology for Graph Neural Networks (GNNs) applied to REM interpolation. Sensors are represented as nodes at their H3 cell centroids, and edges connect neighboring cells. Message-passing layers propagate RF measurements across the graph, enabling spatial-temporal interpolation that respects the true geographic connectivity of the sensor network.

GRID SYSTEM COMPARISON

H3 vs. Geohash vs. S2: Spatial Indexing for REM

Technical comparison of discrete global grid systems for indexing and aggregating radio environment mapping data.

FeatureH3GeohashS2

Cell Shape

Hexagon

Rectangle

Square-like

Hierarchical Resolution Levels

16

12 (variable precision)

30

Area Distortion (max)

0.3%

40%

0.1%

Equal Area Cells

Neighbor Traversal

6 equidistant neighbors

8 neighbors (variable distance)

4-8 neighbors (variable distance)

Parent-Child Containment

Exact hierarchical nesting

Prefix-based truncation

Exact hierarchical nesting

Native Geospatial Operations

Grid distance, hex ring

Bounding box search

Point-in-cell, S2Region operations

Ideal REM Use Case

Uniform propagation modeling, heatmap aggregation

Simple bounding box queries

High-precision signal contouring

H3 HEXAGONAL GRID FAQ

Frequently Asked Questions

Clear, technical answers to the most common questions about Uber's H3 discrete global grid system and its application in radio environment mapping and spectrum awareness.

The H3 hexagonal grid system is a discrete global grid system (DGGS) developed by Uber that partitions the Earth's surface into a hierarchical, multi-resolution array of hexagonal cells. Unlike traditional square grids based on latitude and longitude, H3 uses a gnomonic projection onto an icosahedron, which is then recursively subdivided using an aperture-7 hexagon tiling. This means each parent hexagon contains exactly seven child hexagons at the next finer resolution. The system supports 16 resolution levels, from resolution 0 (122 cells averaging 4,357 km² each) to resolution 15 (569 trillion cells averaging less than 1 m² each). Each cell is identified by a 64-bit integer index that encodes its resolution, base cell, and hierarchical lineage, enabling constant-time spatial queries and neighbor traversal without floating-point geometry calculations.

Prasad Kumkar

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.