AI tools and models fuse AI with geospatial data, science, and technology. They work with a variety of data modalities to automate data creation and extraction, as well as perform complex analysis workflows including classification, pattern detection, and prediction/forecasting.
This chapter covers how AI tools and models are delivered, how models move from pretrained to production, and the architectural decisions that come with managing them over time.
ArcGIS provides AI tools and models through several delivery mechanisms, each suited to different user personas and deployment contexts.
AI-powered geoprocessing tools sit alongside other geoprocessing tools in ArcGIS Pro, ArcGIS Online, and ArcGIS Enterprise. They can be found in toolboxes and chained into workflows just like traditional geoprocessing tools, though under the hood they run ML or DL inferencing. The AI is embedded in the tool providing a simplified experience for GIS analysts. Examples include feature extraction from imagery, point cloud classification, and predictive analysis tools. To learn more, see an Introduction to geoprocessing and the GeoAI toolbox.
AI-powered services in ArcGIS Online provide text analysis and image analysis capabilities. These are accessed through ArcGIS applications like Survey123 and Instant Apps, as well as through the ArcGIS REST API and the ArcGIS API for Python (arcgis.ai module). They do not require customers to manage model infrastructure, as Esri operates the services.
Pretrained models in Living Atlas are published as Deep Learning Packages (DLPKs). Each package includes documentation of its intended use, model architecture, training data, performance metrics, input/output specifications, and limitations. Task-specific models, such as building footprint extraction and land cover classification, can be used directly or fine-tuned. Generalizable models like the Segment Anything Model (SAM) adapt to a wide range of tasks without task-specific training. Browse AI models in Living Atlas.
The ArcGIS API for Python exposes two modules central to AI workflows. The arcgis.learn module is the underlying technology Esri uses to build its own AI tools and pretrained models. It is exposed through the ArcGIS API for Python for organizations looking to train deep learning models on their own data using standard frameworks (PyTorch). Additionally, the arcgis.ai module provides access to text and image analysis services from ArcGIS Online.
End-to-end deep learning workflows in ArcGIS Pro, ArcGIS Notebooks, and Deep Learning Studio support the full model lifecycle from data preparation, labeling, training, evaluation, to deployment. These are aimed at data scientists and GIS analysts looking to train custom models rather than use pretrained ones. Deep Learning Studio provides a visual interface for bringing externally trained models into ArcGIS. Learn more in an Introduction to Deep Learning Studio.
ArcGIS supports a spectrum of approaches from off-the-shelf usage to custom models trained from scratch.
Use pretrained models directly. This is the lowest-effort path, and covers many common extraction and classification tasks. If a task-specific model in Living Atlas matches the use case organizations can deploy it without training. The model runs inference on the organization’s data using weights trained by Esri on representative datasets.
Fine-tune pretrained models. This applies when a pretrained model gets close to the desired outcome, but not close enough. For example, a pretrained model trained on urban imagery may not produce accurate enough results in rural areas, or a model detects building footprints but not the specific roof types needed for a specific use case. Fine-tuning adapts the model to the organization’s data. This requires a much smaller sample of labeled training data than would be required to train a new model from scratch and typically converges faster because the model already has general knowledge.
Train custom models. When no suitable pretrained model exists, or the organization’s data and task are unique enough to warrant it, the arcgis.learn module and ArcGIS Pro support end-to-end deep learning workflows. This path requires more data, more compute (typically GPU), and more expertise — but it provides full control over the model architecture, training process, and resulting behavior.
Geospatial foundation models are a newer direction, currently in research. These are models trained on massive geospatial datasets with billions of parameters, designed to be adapted to many downstream tasks. Esri is developing two types of foundation models. The first is a geodemographic foundation model that combines demographics with location data for similarity search, clustering, and regression. The second is a geospatial vision language model that connects satellite imagery with LLMs for natural-language imagery queries. As these models mature, they may shift the default starting point for many workflows from task-specific pretrained models to foundation model adaptation.
Location embeddings are a newer class of AI output in ArcGIS. They are machine-understandable numerical representations of geography that distill complex location data into vectors that AI models can work with directly. Where traditional geospatial data represents places as coordinates and attributes, embeddings represent places as positions in a high-dimensional space where proximity means similarity.
Location embeddings enable analytical patterns that are difficult or impractical with traditional spatial methods. These include rapid similarity search across large geographies (find locations that resemble a target profile), pattern recognition and change detection (identify where things have shifted over time), clustering without predefined boundaries, and spatial regression where location context is a first-class feature.
Esri’s geodemographic location embeddings combine demographic data with other location characteristics into dense vector representations. The USA geodemographic location embeddings dataset is available through Living Atlas and can be used in similarity search tools provided through ArcGIS Pro. These are early capabilities, and the tooling and available datasets will expand; however, the architectural pattern is important to consider early. Embeddings sit between raw data and model inference as a reusable intermediate representation. Organizations that produce and maintain embeddings for their geographies gain a building block that can feed many downstream workflows without re-processing the source data each time.
The lifecycle for AI tools and models has architectural decisions at every stage.
Selection. It is critical to choose the right starting point: use pretrained models, fine-tune existing models, or train new custom models. The decision depends on how well existing models match the task, how much labeled data the organization has, and how much compute and expertise it can invest. For many common geospatial tasks, such as building footprint extraction, road detection, and land cover classification, pretrained models in ArcGIS Living Atlas are the right starting point.
Training and evaluation. When training is required, architects need to plan for GPU compute (cloud or on-premises), storage for training datasets and checkpoints, and an evaluation framework that measures model performance against the business problem, not just abstract accuracy metrics. Geographic representativeness of training data matters, as a model evaluated only on the areas it trained on may underperform elsewhere.
Deployment. Trained models can be deployed as geoprocessing tools for interactive use, as services for application integration, or as batch processes for large-area inference. The deployment pattern shapes compute requirements, latency expectations, and how results flow into downstream systems. Managed services through ArcGIS Online are the lowest-infrastructure path, whereas self-hosted deployment in ArcGIS Enterprise offers more control over compute and data residency.
Monitoring and maintenance. Models degrade over time as the world they represent changes. For example, land cover shifts, new building types appear, and sensor characteristics drift. Architects should plan for periodic evaluation against ground-truth data, retraining when performance drops below acceptable thresholds, and versioning so that prior model states can be recovered or compared. This is the domain of MLOps, operational practices for managing models in production.
AI tools and models work across all ArcGIS deployment patterns — SaaS, cloud, on-premises, and disconnected. They do not depend on the LLM gateway that AI assistants and agentic AI features require. For organizations in regulated, sovereign, or air-gapped environments, this is the AI capability most readily available today. See Integration and interoperability for more on deployment considerations.