# GeoAI

> Source: https://aiwiki.ai/wiki/geography
> Updated: 2026-07-07
> Categories: AI for Science, Artificial Intelligence
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution to "AI Wiki (aiwiki.ai)".

**GeoAI**, short for geospatial artificial intelligence, is the application of [artificial intelligence](/wiki/artificial_intelligence), especially [machine learning](/wiki/machine_learning) and [deep learning](/wiki/deep_learning), to geographic data and spatial problems. It sits at the intersection of geographic information science (GIScience), spatial statistics, [remote sensing](/wiki/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](/wiki/large_language_model) for geographic question answering and reasoning. By the mid-2020s this shift had produced [foundation models](/wiki/foundation_model) such as IBM and NASA's Prithvi and Google DeepMind's AlphaEarth Foundations, which compress petabytes of multi-sensor satellite data into compact learned embeddings, and AI weather models such as DeepMind's GenCast, which its developers reported beat the leading operational physics-based forecasting system on 97.2 percent of 1,320 tested targets.[^17][^20]

## Definition and scope

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]

## Theoretical background

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]

## History

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](/wiki/convolutional_neural_network) 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.

## Application areas

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](/wiki/graph_neural_network) 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]

## Core techniques

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](/wiki/vision_transformer), 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] A more recent example is SatCLIP, a global location encoder introduced by Microsoft Research in 2023, which trains matched location and image encoders by contrastive learning, pairing roughly 100,000 Sentinel-2 image patches with their coordinates in the way that CLIP pairs images with text, and reports improved prediction on nine location-dependent tasks including temperature, population density, and species recognition.[^19] 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]

## What are geospatial foundation models?

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. Several 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] Clay version 1 is a 632 million-parameter vision transformer, of which about 311 million parameters sit in its masked-autoencoder encoder, and version 1.5 was pretrained on roughly 70 million globally distributed image chips sampled to match global land-use and land-cover statistics.[^23]

AlphaEarth Foundations, released by [Google DeepMind](/wiki/google_deepmind) in July 2025 as part of Google Earth AI, takes a different approach by learning a global embedding field rather than task-specific labels. The model assimilates more than 3 billion observations from nine gridded data sources, including optical Sentinel-2 and Landsat imagery, Sentinel-1 and PALSAR-2 radar, GEDI LiDAR, ERA5-Land climate data, and GLO-30 topography, and distils them into annual 10-metre embeddings for Earth's land surface and coastal waters covering 2017 to 2024, with subsequent annual updates. Each location is summarised by just 64 numbers (64 bytes), which the authors describe as "16x less information per-representation (64 bytes) compared to the next-most compact learned method," and they report that these embeddings "reduced error magnitudes by ~23.9% on average" against the next-best methods in their most data-rich evaluation.[^17] The resulting Satellite Embedding dataset is published in Google's Earth Engine catalogue for use in mapping tasks such as land-cover classification, deforestation monitoring, and water-resource management.[^17]

TerraMind, released by IBM and the [European Space Agency](/wiki/european_space_agency) in April 2025, is described by its developers as "the first any-to-any generative, multimodal foundation model for Earth observation." It was pretrained on 500 billion tokens drawn from about 9 million globally distributed, spatiotemporally aligned samples spanning nine Earth observation modalities in the TerraMesh dataset, and it learns at both a token level and a pixel level so that it can generate one modality from another. In an ESA evaluation on the PANGAEA benchmark, IBM and ESA reported that TerraMind outperformed twelve widely used Earth observation foundation models on tasks such as land-cover classification and change detection by 8 percent or more.[^18] In October 2025 the partners released lighter versions designed to run on edge devices, including directly aboard satellites.[^18]

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; 632M parameters; sponsored by Radiant Earth[^13][^23] |
| AlphaEarth Foundations | Foundation model | Google DeepMind embedding-field model; 64-value annual 10 m embeddings, 2017 to 2024, from nine data sources[^17] |
| TerraMind | Foundation model | IBM and ESA any-to-any generative multimodal EO model; 500B training tokens, nine modalities[^18] |
| SatCLIP | Location encoder | Microsoft Research contrastive location-image embeddings from Sentinel-2 (S2-100K)[^19] |
| 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]

## How are AI models used for weather and climate forecasting?

AI weather models are among the most visible and rigorously benchmarked applications of GeoAI, and since 2023 several have matched or surpassed the physics-based numerical weather prediction systems that had dominated forecasting for decades. Rather than solving the equations of atmospheric physics on a supercomputer, these models learn statistical dynamics directly from the ECMWF ERA5 reanalysis archive and then run in seconds to minutes on a single accelerator. DeepMind's GraphCast, described above, established that a single graph-neural-network model could beat the ECMWF High Resolution Forecast on about 90 percent of more than 1,300 targets while producing a 10-day forecast in under a minute.[^9]

Huawei Cloud's Pangu-Weather, published in *Nature* in July 2023, uses a 3D Earth-Specific Transformer trained on 43 years of reanalysis data. According to its authors it was the first AI system to exceed the accuracy of operational numerical methods for lead times from one hour to seven days, and it completes a 24-hour global forecast in about 1.4 seconds on a single GPU, roughly 10,000 times faster than a conventional model.[^21] DeepMind's GenCast, published in *Nature* in December 2024, moved from deterministic forecasts to probabilistic ensembles by using a [diffusion model](/wiki/diffusion_model) to generate 50 or more possible weather trajectories at 0.25 degree resolution. DeepMind reported that GenCast outperformed the ECMWF ENS ensemble, the leading operational system, on 97.2 percent of 1,320 verification targets, rising to 99.8 percent at lead times beyond 36 hours, and produced a full 15-day forecast in about eight minutes on a single Cloud TPU v5.[^20]

Microsoft's Aurora, published in *Nature* in 2025, generalises the idea into a foundation model of the Earth system trained on "more than one million hours" of geophysical data that can be fine-tuned for tasks well beyond weather, including air quality, ocean waves, and tropical-cyclone tracks. Its authors report that Aurora beats existing numerical and AI baselines on 91 percent of targets when fine-tuned for medium-range weather, and that it matches or outperforms the CAMS air-quality system on 74 percent of atmospheric-pollution tasks without using any emissions data as input.[^22] A recurring caveat across this literature is that the models are trained on and verified against the same reanalysis data, so their reliability on genuinely unprecedented extremes and their physical consistency remain active research questions.

| Model | Developer | Published | Architecture | Notable reported result |
| --- | --- | --- | --- | --- |
| Pangu-Weather | Huawei Cloud | 2023 | 3D Earth-Specific Transformer | First AI model to beat numerical forecasts for 1 hour to 7 days; 24-hour forecast in about 1.4 s[^21] |
| GraphCast | Google DeepMind | 2023 | Graph neural network | Beat ECMWF HRES on about 90% of 1,300+ targets; 10-day forecast in under a minute[^9] |
| GenCast | Google DeepMind | 2024 | Diffusion ensemble | Beat ECMWF ENS on 97.2% of 1,320 targets; 15-day forecast in about 8 min[^20] |
| Aurora | Microsoft | 2025 | Transformer foundation model | Beat baselines on 91% of medium-range targets; adds air quality, waves, cyclone tracks[^22] |

## Can large language models reason about geography?

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.

## Challenges

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]

## Outlook

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, Clay, AlphaEarth, and TerraMind families.[^7][^13][^17][^18] AI weather and climate models such as GraphCast, GenCast, and Aurora suggest that learned systems can match or exceed established physics-based methods on some forecasting tasks while running far faster.[^9][^20][^22] Commercially the field is growing quickly: the market-research firm MarketsandMarkets estimated the geospatial intelligence and GeoAI market at USD 37.13 billion in 2025 and projected it to reach USD 62.88 billion by 2030, a compound annual growth rate of about 11 percent.[^24] 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]

## See also

- [Geography ChatGPT Plugins](/wiki/geography_chatgpt_plugins)
- [Computer vision](/wiki/computer_vision)
- [Convolutional neural network](/wiki/convolutional_neural_network)
- [Large language model](/wiki/large_language_model)
- [Foundation model](/wiki/foundation_model)
- [Remote sensing](/wiki/remote_sensing)

## References

[^1]: Janowicz, K., Gao, S., McKenzie, G., Hu, Y., & Bhaduri, B. (2020). "GeoAI: spatially explicit artificial intelligence techniques for geographic knowledge discovery and beyond." *International Journal of Geographical Information Science*, 34(4), 625-636. https://www.tandfonline.com/doi/full/10.1080/13658816.2019.1684500 Accessed 2026-05-31.
[^2]: Gao, S. (2021). "Geospatial Artificial Intelligence (GeoAI)." *Oxford Bibliographies in Geography*. DOI: 10.1093/OBO/9780199874002-0228. https://geography.wisc.edu/geods/wp-content/uploads/sites/28/2022/05/2021_OxfordBibliographies_GeoAI.pdf Accessed 2026-05-31.
[^3]: Esri. "What Is GeoAI?" https://www.esri.com/en-us/capabilities/geoai/overview Accessed 2026-05-31.
[^4]: GIS Geography. "What is Tobler's First Law of Geography?" https://gisgeography.com/tobler-first-law-of-geography/ Accessed 2026-05-31.
[^5]: GIS Geography. "MAUP: The Modifiable Areal Unit Problem." https://gisgeography.com/maup-modifiable-areal-unit-problem/ Accessed 2026-05-31.
[^6]: Sumbul, G., Charfuelan, M., Demir, B., & Markl, V. (2019). "BigEarthNet: A Large-Scale Benchmark Archive for Remote Sensing Image Understanding." arXiv:1902.06148. https://arxiv.org/abs/1902.06148 Accessed 2026-05-31.
[^7]: Szwarcman, D., et al. (2024). "Prithvi-EO-2.0: A Versatile Multi-Temporal Foundation Model for Earth Observation Applications." arXiv:2412.02732. https://arxiv.org/abs/2412.02732 Accessed 2026-05-31.
[^8]: Google DeepMind. "Traffic prediction with advanced Graph Neural Networks." https://deepmind.google/discover/blog/traffic-prediction-with-advanced-graph-neural-networks/ Accessed 2026-05-31.
[^9]: Lam, R., Sanchez-Gonzalez, A., et al. (2023). "Learning skillful medium-range global weather forecasting." *Science*, 382(6677), 1416-1421. DOI: 10.1126/science.adi2336. https://www.science.org/doi/10.1126/science.adi2336 Accessed 2026-05-31.
[^10]: Van Etten, A., Lindenbaum, D., & Bacastow, T. M. (2018). "SpaceNet: A Remote Sensing Dataset and Challenge Series." arXiv:1807.01232. https://arxiv.org/abs/1807.01232 Accessed 2026-05-31.
[^11]: NASA Earthdata. "NASA and IBM Openly Release Geospatial AI Foundation Model for NASA Earth Observation Data." https://www.earthdata.nasa.gov/news/nasa-ibm-openly-release-geospatial-ai-foundation-model-nasa-earth-observation-data Accessed 2026-05-31.
[^12]: IBM Research. "IBM and NASA release a new version of Prithvi." https://research.ibm.com/blog/prithvi2-geospatial Accessed 2026-05-31.
[^13]: Clay Foundation. "Clay Foundation Model." https://clay-foundation.github.io/model/ Accessed 2026-05-31.
[^14]: Hochmair, H. H., et al. (2024). "Evaluating Large Language Models on Spatial Tasks: A Multi-Task Benchmarking Study." arXiv:2408.14438. https://arxiv.org/html/2408.14438v4 Accessed 2026-05-31.
[^15]: Ji, Y., & Gao, S. (2025). "Foundation Models for Geospatial Reasoning: Assessing Capabilities of Large Language Models in Understanding Geometries and Topological Spatial Relations." arXiv:2505.17136. https://arxiv.org/pdf/2505.17136 Accessed 2026-05-31.
[^16]: "Distortions in Judged Spatial Relations in Large Language Models." (2024). arXiv:2401.04218. https://arxiv.org/pdf/2401.04218 Accessed 2026-05-31.
[^17]: Brown, C. F., et al. (Google DeepMind) (2025). "AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data." arXiv:2507.22291. https://arxiv.org/abs/2507.22291 ; Google Earth Engine Data Catalog, "Satellite Embedding V1." https://developers.google.com/earth-engine/datasets/catalog/GOOGLE_SATELLITE_EMBEDDING_V1_ANNUAL Accessed 2026-07-08.
[^18]: IBM Research. "TerraMind: the new generative AI model for Earth observation." https://research.ibm.com/blog/terramind-esa-earth-observation-model ; IBM and ESA (2025). "TerraMind: Large-Scale Generative Multimodality for Earth Observation." arXiv:2504.11171. https://arxiv.org/abs/2504.11171 Accessed 2026-07-08.
[^19]: Klemmer, K., Rolf, E., Robinson, C., Mackey, L., & Rußwurm, M. (2025). "SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery." *Proceedings of the AAAI Conference on Artificial Intelligence*; arXiv:2311.17179. Microsoft Research. https://arxiv.org/abs/2311.17179 Accessed 2026-07-08.
[^20]: Google DeepMind (2024). "GenCast predicts weather and the risks of extreme conditions with state-of-the-art accuracy." https://deepmind.google/blog/gencast-predicts-weather-and-the-risks-of-extreme-conditions-with-sota-accuracy/ ; "Probabilistic weather forecasting with machine learning." *Nature* (2024). https://www.nature.com/articles/s41586-024-08252-9 Accessed 2026-07-08.
[^21]: Bi, K., et al. (2023). "Accurate medium-range global weather forecasting with 3D neural networks." *Nature*; arXiv:2211.02556. https://arxiv.org/abs/2211.02556 Accessed 2026-07-08.
[^22]: Bodnar, C., et al. (2025). "A foundation model for the Earth system." *Nature*, 641, 1180-1187. https://www.nature.com/articles/s41586-025-09005-y ; Microsoft Research. "Introducing Aurora: The first large-scale foundation model of the atmosphere." https://www.microsoft.com/en-us/research/blog/introducing-aurora-the-first-large-scale-foundation-model-of-the-atmosphere/ Accessed 2026-07-08.
[^23]: Clay Foundation. "Pretrained Model release v1.5." https://clay-foundation.github.io/model/release-notes/specification.html Accessed 2026-07-08.
[^24]: MarketsandMarkets. "Geospatial Intelligence Market worth $62.88 billion by 2030." https://www.marketsandmarkets.com/PressReleases/geospatial-intelligence.asp Accessed 2026-07-08.