GeoAI
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v2 ยท 2,764 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
No citations yet
Review status
Needs citations
Revision
v2 ยท 2,764 words
Add missing citations, update stale details, or suggest a clearer explanation.
GeoAI, short for geospatial artificial intelligence, is the application of artificial intelligence, especially machine learning and deep learning, to geographic data and spatial problems. It sits at the intersection of geographic information science (GIScience), spatial statistics, remote sensing, and modern AI, and is concerned with capturing, representing, analysing, and reasoning about phenomena that have a location on or near the Earth's surface.[1][2] Typical tasks include classifying land cover from satellite imagery, detecting objects such as buildings or roads, predicting traffic and travel times, interpolating values at unsampled locations, geocoding place names, and answering questions about places. GeoAI is sometimes described as a subfield of spatial data science, and a recurring theme in the academic literature is that geographic data have special properties, such as spatial dependence and spatial heterogeneity, that motivate AI methods which are explicitly aware of space rather than treating location as just another feature.[1][2]
The term gained currency in the geography and GIScience research communities in the late 2010s, alongside the broader deep learning boom, and was popularised in part by an influential 2020 editorial in the International Journal of Geographical Information Science.[1] Since then GeoAI has expanded from task-specific neural networks toward large geospatial foundation models trained on petabytes of Earth observation data, and toward the use of large language models for geographic question answering and reasoning.
In their 2020 editorial, Krzysztof Janowicz, Song Gao, Grant McKenzie, Yingjie Hu, and Budhendra Bhaduri framed GeoAI as a subfield of spatial data science that uses advances in AI techniques and in data culture to support the creation of more intelligent geographic information, as well as methods, systems, and services for tasks such as image classification, object detection, scene segmentation, simulation and interpolation, link prediction, natural-language retrieval and question answering, on-the-fly data integration, and geo-enrichment.[1] In Song Gao's Oxford Bibliographies entry, GeoAI is described as the integration of geospatial studies and AI, and as a study subject aimed at developing intelligent computer programs that mimic human perception, spatial reasoning, and discovery about geographic phenomena and dynamics, with a focus on spatial contexts and roots in geography and GIScience.[2] A practitioner-facing description is offered by Esri, which defines GeoAI as the application of AI fused with geospatial data, science, and technology.[3]
A central idea distinguishing GeoAI from generic machine learning is the notion of a spatially explicit model. Drawing on Michael Goodchild's work, a model is considered spatially explicit if it satisfies tests such as an invariance test, a representation test, a formulation test, and an outcome test, meaning that its results depend on the locations of the objects involved and would change if those locations changed.[2] Spatially explicit models that incorporate spatial context have been shown to outperform conventional non-spatial machine learning on tasks such as place-type image classification and geographic question answering.[2]
GeoAI rests on long-standing principles of geography and spatial analysis. Tobler's First Law of Geography, articulated by Waldo Tobler in 1969, states that everything is related to everything else, but near things are more related than distant things.[4] This law underpins the concepts of spatial dependence and spatial autocorrelation, the tendency for values at nearby locations to resemble one another. Positive spatial autocorrelation is the default condition of most geographic phenomena, which violates the standard statistical assumption that observations are independent and identically distributed, and therefore requires methods that account for it.[4] Researchers have extended this thinking with a proposed Second Law (spatial heterogeneity) and a Third Law of geography, the latter focusing on the similarity of geographic configurations between locations as a basis for spatial prediction.[2]
A second foundational concern is the modifiable areal unit problem (MAUP), a statistical bias that arises when point data are aggregated into areal units. MAUP has two components: a scale effect, in which results change as the size of the aggregation units changes (for example, county versus census tract), and a zone effect, in which results change as the shape of the units changes while their number is held constant.[5] Because administrative boundaries are often drawn arbitrarily, conclusions drawn from aggregated spatial data can be unstable, a caveat that carries over into GeoAI models trained on or evaluated against aggregated geographic data.[5]
The intersection of AI and geographic studies is not new. Historical roots are traced to work such as Terence Smith's 1984 paper on AI and geographical problem solving and Helen Couclelis's 1986 essay on artificial intelligence in geography.[2] During the 1980s and 1990s, geographers and regional scientists experimented with expert systems, artificial neural networks, neurocomputing, genetic programming, fuzzy logic, and other techniques, a strand of work associated with Stan Openshaw, whose 1992 paper on AI tools for spatial modelling and his 1997 book with Christine Openshaw, Artificial Intelligence in Geography, are regarded as milestones.[2] This early period overlapped with the rise of geocomputation and, in the 2000s, ontology and the semantic web for geographic information retrieval.[2]
The modern phase of GeoAI was driven by the deep learning breakthroughs reviewed by Yann LeCun, Yoshua Bengio, and Geoffrey Hinton in 2015, combined with a dramatic increase in the availability of geospatial data from satellites, mobile devices, and volunteered geographic information.[2] Convolutional neural networks made it practical to analyse satellite and aerial imagery at scale, and the GeoAI label was consolidated by community efforts including ACM SIGSPATIAL GeoAI workshops beginning in 2017 and the 2020 IJGIS special issue.[1][2] The most recent phase, from roughly 2023 onward, has been characterised by geospatial foundation models and by the application of general-purpose large language models to spatial tasks.
GeoAI spans a wide range of applications. The table below summarises major areas and representative tasks documented in the literature.
| Application area | Representative tasks |
|---|---|
| Remote-sensing image analysis | Land-cover and land-use classification, crop-type mapping, semantic segmentation of satellite and aerial imagery[2][6] |
| Object detection | Detecting buildings, roads, vehicles, ships, and natural terrain features in overhead imagery[1][2] |
| Change detection | Identifying changes between multi-temporal images, such as deforestation, urban growth, floods, and burn scars[7] |
| Spatial prediction and interpolation | Estimating values at unsampled locations, an AI-based complement to inverse distance weighting and kriging[2] |
| Geocoding and routing | Matching place names to coordinates, and estimating travel times and routes on road networks[2][8] |
| Disaster response and climate monitoring | Flood and wildfire mapping, AI-based weather forecasting, atmospheric-river and cyclone tracking[7][9] |
| Urban analytics | Urban functional-region extraction, land-use inference, and built-environment characterisation[2] |
| Mobility analysis | Trajectory modelling, traffic-flow prediction, and human-movement reconstruction from sparse location data[2][8] |
| Location-based services | Place recommendation and point-of-interest similarity from learned spatial representations[2] |
Remote sensing is the most mature application area. Large benchmark archives have enabled supervised training at scale: BigEarthNet provides hundreds of thousands of Sentinel-2 image patches labelled with land-cover classes derived from the CORINE Land Cover database, and the SpaceNet challenge series, beginning in 2016, provided high-resolution satellite imagery for tasks such as building-footprint and road extraction, alongside the IARPA Functional Map of the World competition.[6][10]
In mobility, graph-based deep learning has reached production scale. Google DeepMind, working with the Google Maps team, used graph neural networks to improve the accuracy of real-time estimated times of arrival, reporting reductions in inaccuracy of more than 50 percent in some cities by modelling the connectivity structure of road networks via aggregated road segments the team called Supersegments.[8]
In climate and disaster work, AI weather models have become a notable example of GeoAI at planetary scale. DeepMind's GraphCast, published in Science in 2023, is an autoregressive graph-neural-network model trained on the ECMWF ERA5 reanalysis archive that produces a 10-day global forecast at 0.25 degree resolution in under a minute, and outperformed the ECMWF High Resolution Forecast on roughly 90 percent of more than 1,300 verification targets according to its authors.[9]
The technical toolkit of GeoAI is largely inherited from mainstream deep learning and adapted to spatial data. Convolutional neural networks remain a workhorse for gridded imagery, learning local spatial features that are well suited to pixels and patches.[2] Vision transformers, which apply the attention mechanism to image patches, underpin many recent Earth observation foundation models.[7] Graph neural networks are used where data are naturally relational, such as road networks, sensor networks, and movement graphs, because they can perform spatiotemporal reasoning over connectivity structure.[8]
A distinctive line of GeoAI research concerns spatial representation learning, which seeks vector embeddings that encode location and spatial context. Examples documented in the literature include Place2Vec and related methods for place-type embeddings, Road2Vec for road segments and traffic prediction, Mot2Vec for mobility traces, Tile2Vec for remote-sensing tiles, and the multi-scale Space2Vec location encoder for spatial feature distributions.[2] Generative methods have also been explored, including GeoGAN for generating map layers from satellite images and SpaceGAN, which augments correlation structure in spatial data.[2] Throughout, the methodological literature emphasises that ignoring spatial autocorrelation and MAUP, or evaluating models without spatially aware cross-validation, can produce misleadingly optimistic results.[2][5]
A major development since 2023 is the geospatial foundation model (GFM), a large model pretrained in a self-supervised way on vast amounts of Earth observation imagery and then fine-tuned for downstream tasks. Two openly released families illustrate the trend.
Prithvi, developed by IBM and NASA, was released in 2023 as a temporal vision transformer pretrained on NASA's Harmonized Landsat and Sentinel-2 (HLS) data at 30-metre resolution; the initial Prithvi-100M model has about 100 million parameters and was published openly on Hugging Face under NASA's open-science initiative for tasks including flood mapping, burn-scar and land-cover mapping, and crop-type classification.[7][11] The second generation, Prithvi-EO-2.0, was developed by IBM, NASA, and the Julich Supercomputing Centre with collaborators across multiple institutions, and was released in 300-million-parameter (ViT-L) and 600-million-parameter (ViT-H) variants pretrained with a masked-autoencoder approach on 4.2 million global time-series samples from the HLS archive, with temporal and location embeddings. Its authors report that the 600M model outperforms the previous Prithvi-EO model by about 8 percent across a range of tasks on the GEO-Bench benchmark and outperforms six other geospatial foundation models across resolutions from 0.1 to 15 metres.[7][12] Prithvi is distributed through Hugging Face and IBM's TerraTorch toolkit.[12]
Clay is an open-source Earth foundation model sponsored by the non-profit Radiant Earth Foundation. It uses a vision-transformer architecture trained by self-supervised learning with a masked-autoencoder method, and takes satellite imagery together with location and time as input to produce embeddings, numerical representations of a place at a moment in time, that can be used for similarity search or fine-tuned for classification, regression, and other tasks. Clay's code, model weights, and training-data embeddings are released openly under permissive licenses on GitHub, Hugging Face, and Source Cooperative.[13]
The table below lists notable models, datasets, and tools that are publicly documented.
| Name | Type | Notes |
|---|---|---|
| Prithvi-EO / Prithvi-EO-2.0 | Foundation model | IBM and NASA temporal vision transformer on HLS data; 100M, 300M, and 600M variants[7][11][12] |
| Clay | Foundation model | Open vision-transformer Earth model producing embeddings; sponsored by Radiant Earth[13] |
| GraphCast | AI weather model | DeepMind graph-neural-network model trained on ERA5; 10-day global forecasts[9] |
| BigEarthNet | Dataset | Hundreds of thousands of labelled Sentinel-2 patches (CORINE land-cover labels)[6] |
| SpaceNet / Functional Map of the World | Datasets/challenges | High-resolution imagery for building, road, and functional-area extraction[6][10] |
| ArcGIS GeoAI tooling | Software | Esri deep-learning tools, pretrained models, and integration with PyTorch-based libraries[3] |
Commercial GIS platforms have also incorporated GeoAI. Esri's ArcGIS provides deep-learning tools for extraction, classification, and detection from imagery, video, point clouds, and text, a library of pretrained models that require no training data, and integration with open-source packages such as MMDetection and MMSegmentation through the ArcGIS Pro Image Analyst module.[3]
The rise of large language models has prompted research into whether they can reason about space and geography. Studies have probed tasks such as inferring topological relations between geometries, judging relative directions and distances between places, and planning paths. The results are mixed and generally show meaningful limitations. One evaluation found that GPT-4 reached about 55 percent accuracy on spatial-relation tasks overall, dropping to roughly 33 percent on items suspected of hierarchical bias, while performing better, around 86 percent, on others.[14] Other work reports that larger GPT models can map natural-language text to spatial relations reasonably well but struggle with multi-hop spatial reasoning, and that even advanced models face difficulty producing consistent solutions in grid-based spatial-planning benchmarks.[15] A separate analysis documented systematic distortions in the spatial relations that language models judge between real-world places, attributing them partly to biases in training text.[16]
These findings inform a broader pattern: language models can serve as natural-language interfaces to geographic data and can sometimes answer simple locational questions, but they are unreliable for precise geometric or topological computation and are prone to spatial bias. A common architectural response is to pair an LLM with authoritative geographic services or tools rather than relying on its parametric knowledge, an approach exemplified historically by map and routing plugins.
GeoAI faces several recurring challenges. Spatial bias and data quality are persistent concerns: training data are unevenly distributed across the globe, with far richer coverage of wealthy, well-mapped regions, which can cause models to perform worse in under-represented areas. The statistical peculiarities of geographic data, spatial autocorrelation and the MAUP, can inflate apparent accuracy if models are not evaluated with spatially aware validation.[2][5] Privacy is a significant issue for mobility and trajectory data, motivating privacy-preserving techniques such as synthetic-trajectory generation.[2] Interpretability is also emphasised, given that many GeoAI models are opaque while the decisions they inform, in urban planning, disaster response, and resource management, can have real consequences. Finally, reproducibility depends on the open sharing of high-quality geospatial datasets, a need explicitly highlighted by the GeoAI research community.[1]
The trajectory of GeoAI points toward larger and more general geospatial foundation models, increasingly trained on multi-sensor and multi-temporal Earth observation data and released openly, as exemplified by the Prithvi and Clay families.[7][13] AI weather and climate models such as GraphCast suggest that learned systems can match or exceed established physics-based methods on some forecasting tasks while running far faster.[9] At the same time, integrating language models with rigorous spatial computation, and addressing spatial bias, privacy, and interpretability, remain open problems. The longer-term ambition expressed in the founding GeoAI literature, to build spatially explicit AI that genuinely advances geographic knowledge discovery rather than merely applying off-the-shelf models to maps, continues to shape the field.[1][2]