Geo: Geographic, Spatial & Temporal Computing

• Geo is a multidisciplinary domain that explicitly represents location and time through coordinates, polygons, and spatial relations in computational models.
• It advances joint representation learning by integrating geometric features with language models and knowledge graphs to enhance spatial reasoning and retrieval.
• Geo bridges traditional GIS and modern deep learning, supporting applications from geospatial analytics to agentic systems and operational platforms.

Geo, in contemporary arXiv literature, denotes a family of geographic, geospatial, and often geo-temporal representations, models, and systems whose common requirement is explicit treatment of where phenomena occur and, in many settings, when they occur. In this usage, geographic space is not incidental metadata but a first-class computational object: it is represented through coordinates, polygons, administrative units, spatial relations, mobility patterns, or geostationary orbital geometry, and is coupled with text, knowledge graphs, retrieval pipelines, or generative models to support reasoning, prediction, and synthesis (Martins et al., 17 Jun 2025, Jeansoulin, 2015, Hu et al., 2024). The term is also polysemous: in some subfields, “GEO” names specific constructs such as the Geography and Election Outcome metric, Generative Engine Optimization, or geostationary Earth orbit network models (Campisi et al., 2021, Aggarwal et al., 2023, Jung et al., 2023).

1. Semantic range and historical formation

Within geo-information science, “geo” historically referred to remote sensing, automated mapping, photogrammetric engineering, and geographic information systems. One retrospective account states that geo-information “pioneered what today is termed ‘big data’,” noting that Landsat “was already delivering gigabytes of data” in 1983 and that the field had already been forced to confront structure, metadata, processing, uncertainty, consistency, and ontologies at scale (Jeansoulin, 2015). The same literature distinguishes raster data as sampled signals on grids and vector data as points, lines, and polygons, and emphasizes that topology was explicitly encoded so that relations such as contains, is in, touches, and has hole could be reasoned about without recomputing geometry each time (Jeansoulin, 2015).

This historical trajectory matters because many current uses of “Geo” inherit exactly these commitments. Metadata evolved from “ancillary data” to a central representational layer; standards such as ISO 19101, ISO 19113, ISO 19150, and ISO 19157 codified conceptual schemas, ontologies, and quality models; and geo-information engineering developed an explicit vocabulary for completeness, conceptual consistency, topological consistency, positional accuracy, temporal consistency, and thematic accuracy (Jeansoulin, 2015). A plausible implication is that current GeoAI systems are extensions of this earlier schema-centric tradition rather than a break from it.

The same historical literature also insists that geo-data are never merely geometric. They are entwined with uncertainty, semantic alignment, and decision making. Its formulation that “Data are not Facts, but Acts” captures a core geographic stance: spatial representations are products of measurement, modeling, and ontology, not transparent copies of the world (Jeansoulin, 2015).

2. Representation, topology, and geographic knowledge structures

Recent work makes “Geo” explicit at the level of representation learning and knowledge modeling. In geospatial knowledge graphs, facts are represented as triples $(h, r, t)$, but the geo-specific extension adds geometries for entities and spatial relation semantics such as topology, direction, and distance (Hu et al., 2024). A geometry-enhanced KGE model built on HAKE aligns relation-term embeddings with geometric feature embeddings derived from the 9-intersection model, an 8-point compass, and 20 Jenks natural-break distance categories; on its GeoKG, overall MRR rises from 0.189 for HAKE to 0.208 with topology+direction and 0.207 with topology+direction+distance, while geographic entity prediction reaches 0.215 MRR with all three feature types (Hu et al., 2024). This establishes a specifically geographic notion of plausibility: “adjacent to” should not retrieve geographically distant entities merely because of symbolic co-occurrence.

A more ontological formulation appears in “Geo-Situation,” which specializes dolce:Situation for geographic knowledge graphs (Stephen et al., 2022). A Geo-Situation is a structured spatio-temporal snapshot involving geo-objects and/or geo-events plus observations, and it is used to answer “why” questions about hazards and other geographic events. The framework distinguishes GeoObject, GeoEvent, and GeoSituation, and defines three causal relations: causes for event–event causation, effects for situation–event precondition relations, and affects for situation–object influence (Stephen et al., 2022). It is implemented conceptually by combining DOLCE-Lite, GeoSPARQL, OWL-Time, and O&M, so that geometry, time, and observation semantics all participate in causal explanation (Stephen et al., 2022).

A parallel line of work treats “Geo” as a problem of unified neural representation for points, polylines, and polygons. Geo2Vec learns signed distance fields directly in the original coordinate space, adaptively samples near vertices and edges, and uses a rotation-invariant positional encoding to build a shared representation space for heterogeneous geo-entities (Chu et al., 26 Aug 2025). On the Building dataset, it reports shape accuracy 97.34 and edge MAE 2.22; on Singapore, it reports line-length MAE 5.75 and polygon–polygon distance MAE 5.5, outperforming Poly2Vec and related baselines on shape, topology, and distance tasks (Chu et al., 26 Aug 2025). Together, these strands indicate that “Geo” increasingly denotes a representational regime in which geometry, relation structure, and spatial semantics are learned jointly.

3. Geo in LLMs, retrieval, and deep research

In LLM-centered systems, “Geo” has become closely associated with explicit spatial and temporal grounding. “Geo-temporal deep research systems” are defined as agentic, LLM-driven research pipelines that perform iterative search, retrieval, and reasoning with explicit awareness of geographic and temporal constraints, so that questions tied to where and when can be answered without treating place and time as mere keywords (Martins et al., 17 Jun 2025). The proposed architecture includes query understanding and planning, geo-temporal retrieval and re-ranking using GIR and TIR methods, summarization with geo-coders, reverse geo-coders, gazetteers, and lightweight GIS operations, and reinforcement learning over the pipeline; locations may be represented as points, polygons, administrative units, or hierarchies, and time as points, intervals, months, quarters, or years (Martins et al., 17 Jun 2025). The same vision is explicit that such systems are not intended to replace full GIS, but to provide long-form, text-based synthesis that is geo-temporally grounded and can be plugged into downstream GIS or data analysis workflows (Martins et al., 17 Jun 2025).

This shift has produced dedicated geographic NLP benchmarks. GeoGLUE defines six tasks—GeoTES-recall, GeoTES-rerank, GeoETA, GeoCPA, GeoWWC, and GeoEAG—and evaluates Chinese PLM baselines with metrics tailored to retrieval, sequence tagging, and entity alignment (Li et al., 2023). On the benchmark, StructBERT reaches 43.10 MRR@5 on GeoTES-recall, 83.51 MRR@1 on GeoTES-rerank, and Nezha reaches 79.77 Macro-F1 on GeoEAG, while GeoCPA and GeoWWC remain substantially harder, with best Micro-F1 scores of 67.70 and 70.13 respectively (Li et al., 2023). The benchmark therefore operationalizes geographic language understanding as a combination of POI search, address parsing, where–what segmentation, and entity alignment.

Task-specific models then exploit this structure. Geo-Encoder introduces a chunk-argument bi-encoder for Chinese geographic reranking, using MGEO tagging to identify geographic chunks and an asynchronous update mechanism for chunk-type attention; on GeoTES, it raises the Hit@1 score of MGEO-BERT by 6.22%, from 62.76 to 68.98 (Cao et al., 2023). ERNIE-GeoL goes further by treating geography as a pre-training substrate: it uses a heterogeneous graph with POI and query nodes, Query-click-POI, Origin-to-Destination, and POI-(co-locate with)-POI edges, plus a geocoding objective over S2 cells, and has been deployed in production at Baidu Maps since April 2021 (Ren et al., 2022). Its gains are largest where spatial grounding is decisive: geocoding Accuracy@3km rises to 0.6545, compared with 0.4636 for ERNIE 2.0, and address parsing reaches 0.8794 $F_1$ (Ren et al., 2022).

4. Agentic geo-systems and operational platforms

The same geo-specific grounding appears in operational systems that combine language interfaces with external tools. Geode is a zero-shot geospatial question-answering agent that uses an LLM as a code-generating planner over a pool of geospatial experts, rather than answering directly from parametric memory (Gupta et al., 2024). Its core abstraction is GeoPatch, which stores raster data and vector data, while expert APIs perform geocoding through Nominatim, meteorological and air-quality retrieval through WeatherAPI.com and OpenMeteo, interpolation via radial basis functions, and operations such as thresholding, intersection, and correlation (Gupta et al., 2024). The system is designed to answer questions such as “Where does it rain more, Atlanta or Chicago?” or “What is the air quality like in the city known for the Qutub Minar?” with a textual answer plus a map visualization (Gupta et al., 2024).

Georacle addresses a different systems setting: geospatially aware smart contracts (Azzaoui, 2021). Built on Chainlink and OpenStreetMap, it provides functions such as nodesInArea, nodeCountInArea, nodesInBB, wayGeometry, geocode, and reverseGeocode, while keeping on-chain representations compact through 64-bit identifiers and coordinates scaled by a factor of $10^8$ (Azzaoui, 2021). The design is explicitly constrained by blockchain storage, gas consumption, and lack of floating point arithmetic, so geo-operations are pushed off-chain and only typed results are returned on-chain (Azzaoui, 2021).

Cross-view geo-spatial learning supplies yet another operational meaning. Geo$^2$ jointly addresses Cross-View Geo-Localization and bidirectional Cross-View Image Synthesis by injecting 3D geometric priors from VGGT into a shared latent space through GeoMap and into a flow-matching generator through GeoFlow (Zhang et al., 26 Mar 2026). On CVACT Val it reports R@1 = 94.36 for localization, and on CVACT it reports G2S FID 31.72 and S2G FID 27.77, showing that a common geometry-aware latent space can support both retrieval and generation (Zhang et al., 26 Mar 2026). In a different multimodal direction, SounDiT defines Geo-Contextual Soundscape-to-Landscape generation, introduces the SoundingSVI dataset with 334,495 soundscape–landscape pairs across 90 countries and the SonicUrban dataset with 250,310 pairs across 131 cities and 97 countries, and evaluates generated images with a Place Similarity Score over environmental elements, scene categories, and human perception (Wang et al., 19 May 2025).

5. GIS integration, networked space, and orbital Geo

Some uses of “Geo” remain explicitly GIS-centered. “(geo)graphs” defines a graph $G=(V,E)$ whose nodes have known geographical locations and whose edges have spatial dependence, and operationalizes this definition as a point shapefile for nodes plus a line shapefile for edges (Santos et al., 2017). The framework is realized by the open-source tools GIS4GRAPH and GeoCNet and is applied to street networks in Lorena/SP, origin–destination flows between traffic zones in Rio de Janeiro, and rainfall-correlation networks near Nova Friburgo (Santos et al., 2017). The contribution is not a new spatial graph theory so much as a bridge between complex-network analytics and GIS/GDBMS environments, allowing degree, betweenness, clustering, and shortest paths to be mapped back onto georeferenced layers (Santos et al., 2017).

In satellite communications, “GEO” narrows to geostationary Earth orbit. A stochastic-geometry model of GEO satellite networks places satellites on the geostationary orbit according to a binomial point process, studies visibility as a function of terminal latitude, and derives distance distributions and coverage probability (Jung et al., 2023). The visible arc length depends on latitude, and the limit latitude beyond which the visible arc vanishes is $$\phi_{\text{max}} = \cos^{-1}\!\left(\frac{r_e}{r_e+a}\right) \approx 81.3^\circ$$ (Jung et al., 2023). The paper then derives case probabilities for zero, one, or more than one visible satellite, obtains serving and interfering distance distributions, and approximates the BPP analysis through the Poisson limit theorem; Monte Carlo simulations and TLE-based comparisons show strong alignment with the analytical model (Jung et al., 2023). Here “Geo” is not a general spatial AI term but a specific orbital regime whose geometry controls visibility, interference, and coverage.

6. Acronymal extensions, evaluation regimes, and open problems

The literature also contains acronymal uses in which GEO is not a geographic framework in general but a named metric or optimization paradigm. The “Geography and Election Outcome (GEO) metric” diagnoses partisan gerrymandering by combining district adjacency with district-level vote shares and returns, for each party, a non-negative integer counting previously lost districts that could plausibly be made 50%–50% without risking current wins through local, geographically reasonable changes (Campisi et al., 2021). The metric explicitly depends on the districting graph and regional averages $A_i$, and is presented as a bridge between purely geometric shape metrics and purely election-result metrics (Campisi et al., 2021).

“GEO: Generative Engine Optimization” defines yet another expansion, this time outside geographic computing (Aggarwal et al., 2023). It formalizes visibility in generative search engines through impression metrics such as Position-Adjusted Word Count and Subjective Impression, introduces GEO-bench with 10,000 queries, and shows that content-editing strategies such as quotation addition, statistics addition, and cite sources can boost visibility by up to 40% in generative engine responses (Aggarwal et al., 2023). This use is terminological rather than spatial, but it underscores the breadth of the acronym.

Across the genuinely geographic strands, the major open problems are convergent. Geo-temporal deep research requires open, large-scale search infrastructures, stable shared corpora and index versions, geo-temporal ranking, integration of geo-coding and geo-spatial operations, and multi-dimensional evaluation beyond NDCG-like metrics (Martins et al., 17 Jun 2025). Geo-representation learning continues to seek better joint modeling of topic, spatial relations, and temporal relations (Martins et al., 17 Jun 2025, Hu et al., 2024). Multimodal geo-systems still confront coverage bias, sparse observations, and the need for interoperable tools that can link text, geometry, raster data, and human behavior (Gupta et al., 2024, Wang et al., 19 May 2025). This suggests that “Geo” is increasingly less a domain prefix and more a systems principle: location, geometry, and often time must be represented explicitly, grounded operationally, and evaluated with respect to real spatial structure rather than treated as incidental attributes.