Data is the foundation of every AI system. The accuracy, structure, and accessibility of data directly shape what AI can perceive, reason about, and produce. For geospatial AI in ArcGIS, that foundation is geospatial data, either produced and managed within an organization, built by authoritative providers, or sourced from the community of geospatial organizations and researchers around the world. This chapter covers the data modalities, information models, and quality expectations that AI workloads introduce, and the spatial-specific considerations that come with them.
Geospatial AI in ArcGIS works across several data modalities, each with its own characteristics for training and inference. Each is used differently depending on the AI workload: to train models, to ground inference, and to be discovered, queried, and reasoned against through tools. The latter centers around agentic AI workflows and places particular weight on metadata, schemas, and semantic clarity. Note that these modalities are useful categorizations but not definitive boundaries, many data types and workflows fuse data from across these different modalities or use hybrid approaches.
Vector data — feature layers with points, lines, and polygons are the backbone of most GIS systems. For AI tools and models, vector data serves as both training input (labeled features for supervised learning) and inference output (new features generated by a model). For assistants and agents, vector data is queried, filtered, and reasoned against through tools. A natural-language question becomes a spatial query against feature layers, returning the features that answer it. The attribute schemas, topology, and spatial relationships encoded in vector data give all three modes structured context that unstructured formats cannot. In many cases, tabular data without a location column also falls into this category and is integrated into data models and storage options that also apply to vector data.
Imagery — satellite, aerial, drone, and street-level imagery are the foundation for computer vision workflows in ArcGIS, including object detection, classification, segmentation, and change detection. Training requires annotated imagery with known labels; inference runs on new imagery at scale. The spatial resolution, spectral bands, and temporal coverage of the imagery all affect what a model can learn and how well it generalizes. Assistants and agents can also invoke imagery analysis tools and retrieve imagery for context. Learn more about imagery and remote sensing in ArcGIS.
3D data — point clouds, 3D meshes, building models, and scene layers carry spatial structure that two-dimensional representations lose. AI workflows on 3D data include feature extraction from LiDAR, building footprint generation, classification of point clouds, volumetric analysis, and computer vision patterns applied to snapshots of 3D scenes or content. Models trained on 3D modalities need different architectures than those trained on flat imagery, and the data volumes involved make compute planning more significant. Learn more about 3D GIS.
Text and documents — ArcGIS carries text in many forms: attribute text in features and tables (inspection notes, work orders, maintenance logs), document attachments and items (PDFs, reports), survey responses, and metadata. For AI tools and models, ArcGIS supports a broad range of natural language processing (NLP) including classification, entity extraction, summarization, translation, question answering, and sentiment analysis. These are provided through text analysis tools, the arcgis.learn.text module of the ArcGIS API for Python, AI-powered services in ArcGIS Online, and pre-trained NLP models in Living Atlas. For assistants and agents, text serves as grounding material and as the medium of interaction itself.
Knowledge graphs — graph-structured representations of entities and the relationships between them, modeling real-world systems that bring together spatial and non-spatial data. ArcGIS Knowledge stores knowledge graphs in a dedicated graph store and supports editing, querying (including the openCypher query language), and link analysis. For AI tools and models, knowledge graphs support pattern detection and investigations that can power fraud, intelligence, and complex relationship discovery workflows. For assistants and agents, knowledge graphs provide exceptionally strong grounding. The semantic structure (entities, relationships, properties, ontology) gives agents a reliable scaffolding to reason against that tends to be more accurate than unstructured retrieval. The ArcGIS Pro Assistant supports the AI-assisted conversion of natural language questions into openCypher queries for use with knowledge graphs. Learn more about ArcGIS Knowledge.
Real-time and time series — sensor feeds, IoT telemetry, real-time streams, and temporal feature classes capture change over time. For AI tools and models, forecasting, anomaly detection, and trend analysis operate on time-stamped spatial data. For assistants and agents, real-time and historical telemetry is queried for current state, recent changes, and patterns — an agent monitoring infrastructure or responding to events draws on these sources through tools. Real-time streaming and historical time series differ architecturally in latency, throughput, and windowing, but both are increasingly central to AI workloads particularly for monitoring, prediction, and event-driven workflows. Learn more about real-time data in ArcGIS.
These modalities are not always separate in practice. Multi-modal AI, or models that combine imagery with text, or vector data with temporal signals, is an active area of development. Esri’s geospatial vision language model research, for example, connects satellite imagery with LLMs to allow natural-language queries against image content. As these capabilities mature, the boundaries between modalities will matter less in design than the quality and accessibility of the underlying data.
ArcGIS structures geospatial data through information models that carry semantics beyond raw geometry and pixels. These provide additional grounding to AI.
One information model of note is the geodatabase. Geodatabases encode rules and relationships that enable rich geospatial capabilities, such as topologies and networks. Indoor spaces, buildings, and imagery mosaics are other examples of general purpose information models used widely in geospatial workflows with ArcGIS.
Some information models are industry specific, such as the utility network, parcel fabric, pipeline referencing, and roads and highways models. The domain-specific rules and relationships inherent in these models make them particularly valuable to AI assistants and agents. An agent grounded on a utility network model sees not only the geospatial features, it understands connectivity, flow direction, containment, non-spatial assets, and the rules that govern valid network states.
Geospatial information models tend to outperform unstructured document retrieval for geospatial AI workflows. When an assistant or agent can draw on defined schemas, typed relationships, and governed content, its outputs are more reliable than when it must infer structure from flat text. These models are complementary to, not a replacement for, unstructured and traditional, relational data. From an architectural perspective, the incorporation of geospatial information models should be considered as a first-class design decision.
AI workflows need reliable, trusted data. In ArcGIS, authoritative geospatial content comes from three sources.
Esri-provided content through ArcGIS Living Atlas of the World includes foundational basemaps, boundary datasets, demographic layers, environmental data, imagery, elevation data, and more. Living Atlas also hosts pre-trained models, including task-specific models (e.g., building footprint extraction, land cover classification with Sentinel-2 imagery) that can be fine-tuned to an organization’s data, as well as generalizable models like the Segment Anything Model (SAM) that adapt to a wide range of tasks without specific training. Browse Living Atlas.
Third-party content from trusted contributors and partners brings domain-specific data that Esri does not create directly, including data from national, state and local agencies, weather feeds, commercial imagery, industry-specific datasets, and real-time sensor networks. These sources can serve as training data, as grounding context, or as inference inputs depending on the workflow. Explore data partners.
Organizational content — an organization’s own geospatial data, including systems of record that are validated and maintained by their teams. These are often the most valuable input for AI. An organization’s asset registry, inspection history, network model, or field survey data carry specificity, relevance, and authority that general-purpose content cannot match. These datasets are uniquely available to AI workloads that are run by this organization, and can bring deep meaning and insights when added to other authoritative sources. This content is also where readiness gaps are most common. Learn more about data management in ArcGIS.
Effective AI depends on data that is accurate, well-documented, and governed. This is not a new principle, but AI raises the stakes. A model trained on mislabeled imagery will confidently produce wrong answers. An agent grounded on stale or poorly described content will surface irrelevant results. The failure modes are not always obvious because AI outputs can look plausible even when the underlying data is flawed.
What “AI-ready” means in practice:
Most organizations have data that is of high value to AI workflows, but that is not yet ready to support them in practice. This gap is usually related to data quality and readiness, and closing that gap is often the highest-leverage preparatory work an organization can do before deploying AI capabilities.
AI agents and models need not only access to data, but the right kind of access delivered through governance patterns that ensure security, traceability, and appropriate use. Data in ArcGIS is exposed through multiple patterns, each with different implications for agentic systems.
Web services and data storage — Geospatial data can be accessed through web services that apply business logic and access controls at query time, or as raw data stored in data lakes, databases, or file systems. Web services provide a controlled interface where governance rules are enforced consistently; agents query features through feature services or invoke analysis services that apply predefined workflows. Direct access to stored data offers flexibility but requires agents to be deeply integrated into infrastructure and governance. Most mature AI deployments use a hybrid approach, with agents querying through services for trusted datasets and accessing raw data directly only where necessary.
Authorization and least-privilege access — An agent’s data access should be scoped to what it needs for its assigned task. Just as a human user should not have access to all organizational data, an agent should have access only to the data layers it requires to fulfill its function. This principle, combined with service-based access patterns, prevents accidental or malicious exposure of sensitive information. As agents become embedded in organizational processes, authorization patterns become as critical as they are for human users.
Governance and auditability — Every query an agent executes, every dataset it accesses, and every inference it makes should be traceable. Audit logs that record which agent accessed which data, when, and for what purpose are essential for debugging, for compliance, and for building organizational trust in AI-assisted decisions. Governance frameworks should define what data an agent can access, log all interactions, and provide investigators the trail to understand how a particular recommendation or output was derived.
Data creation and update — As agentic systems mature, agents may graduate from read-only access to creating or editing data — submitting new observations, flagging updates, or triggering data validation workflows. This introduces new access considerations: agents that can create or modify data must be subject to even stricter authorization and auditing. The lineage of agent-created data, the verification workflows that validate it, and the rollback mechanisms if an agent introduces errors all become operational requirements. Data governance must evolve to encompass not just discovery and read access, but creation and change management as well.
Location adds dimensions to AI workloads that generic AI guidance does not address. The following are some considerations unique to spatial data.
Location as first-class context. Most AI systems have no inherent sense of geography. They cannot distinguish between two similar addresses in different cities, reason about proximity, or account for the fact that spatial relationships shape outcomes. Geospatial data brings forms of context, including containment, adjacency, distance, connectivity, temporal-spatial patterns, that AI systems can use but cannot derive without purpose-built tools and data. This is a fundamental value proposition of combining GIS with AI.
Temporal dimensions. Many geospatial datasets carry a time component: real-time sensor feeds, time-series imagery, historical feature snapshots, temporal feature classes. AI workloads that operate on this data need to handle temporality explicitly, including time-aware training, temporal windows for inference, and awareness of how patterns evolve. Temporal alignment across time zones, while straightforward in concept, also tends to be a common source of error.
Location embeddings. Location embeddings are spatial-specific, numerical representations of geography that distill complex location data into forms AI models can incorporate efficiently. Embeddings can be applied as input to downstream AI workflows, used to find similarity across locations, or combined with other features to enhance models. Location embeddings represent a new class of geospatial data product worth treating as architecturally first-class alongside raw data. Location embeddings are discussed further in AI tools and models in the context of how they are produced and applied.
Coordinate reference systems and scale. Data collected at different scales, in different projections, or at different resolutions requires careful handling before it can be combined for AI training or inference. A model trained on 30cm imagery will not generalize well to 3m imagery without adaptation. These are familiar GIS concerns, but they carry new weight when data feeds automated pipelines that do not have a human checking each input.
Spatial bias and representativeness. Training data that over-represents urban areas, particular geographies, or certain sensor types will produce models that underperform elsewhere. Architects should evaluate geographic coverage and representativeness as part of model evaluation, not just overall accuracy metrics.
Spatial privacy. Location data can reveal sensitive information about individuals, assets, or operations even when other identifiers are removed. AI workflows that ingest, process, or output spatial data should account for privacy implications — particularly when data crosses organizational boundaries or feeds agents with broad access.