AI-Enabled Map Production Using Microsoft's MARS Model
Maps have long served as the foundation for how organizations understand, navigate, and make decisions about the world. Yet generating and maintaining maps remains a highly manual, time-consuming, and costly process with susceptibility to human error. Traditional mapping workflows rely on skilled cartographers to manually analyze satellite imagery, identify geographic features with a human eye, and digitize them into map layers. The challenge of keeping maps up to date is well said by Michael Harres, SVP of Engineering Maps at TomTom. “Traditional maps, including early digital versions, were static snapshots, reliant on manual updates, outdated and unable to reflect real-time changes.” - Michael Harres, SVP of Engineering Maps at TomTom (Source: From Paper To AI: The Evolution Of Maps And What’s Next) While recent advances in generative AI have demonstrated impressive visual reasoning capabilities, most general-purpose models are not designed to generate the structured vector outputs (e.g. points, lines, and polygons) required by GIS workflows. They often produce rasterized sketches, illustrations, or incomplete representations that cannot be directly incorporated into enterprise mapping systems. ple Output from a General-Purpose AI Model Prompted to Generate a Map Today, we are excited to release the Map Autoregressive (MARS) model on Microsoft Foundry! MARS is a Geospatial AI model that automatically converts satellite imagery directly into structured, GIS-ready vector map data. Unlike traditional mapping approaches that rely on complex multi-stage pipelines, heuristic processes, and extensive post-processing, MARS’s key differentiator is its unified end-to-end architecture that automatically extracts geographic features and generates vector representations suitable for mapping and geospatial analysis. Unlike segmentation masks, vector outputs can be directly loaded into GIS systems, edited by analysts, and incorporated into downstream mapping, analytics, and operational workflows. Trained on large-scale geospatial datasets, MARS currently supports feature extraction for buildings, roads, railways, and waterbodies with a scalable ontology designed to expand over time. Microsoft recently published a research paper on MARS (http://aka.ms/marspaper), accepted by the International Conference on Learning Representations (ICLR) Conference, that provides a deep technical overview of the model’s architecture, training methodology, and performance. The MARS model from the paper uses the same architecture, adapted through a separate training process, and attains state-of-the-art performance on public benchmarks. Metrics for the Foundry model are available on its model card, which also demonstrates its ability to generate GIS-ready vector representations directly from satellite imagery. Figure 2: Microsoft MARS Model – This image shows a feature extraction output of MARS overlaid on input aerial imagery of London. The extracted features are rendered directly on top of the source image to illustrate how the model converts raw imagery into structured geospatial data. Buildings, roads, railways, and waterbodies represented as vector geometries, enabling organizations to generate and update maps at scale without labor-intensive manual digitization. Legend: Buildings → Orange polygons Roads → Yellow lines Railways → Red lines Waterbodies → Blue shaded polygons The Value of Automated Feature Extraction using MARS The volume of Earth observation data, and the pace at which infrastructure is changing around the world, is growing faster than organizations can manually track. As a result, there is an increasing need for automated solutions that can transform imagery into up-to-date maps at scale. With the MARS model, organizations can: Accelerate map production workflows with vectorized GIS outputs of map features Reduce manual human-labeling efforts Scale mapping operations across larger regions or under-represented areas Increase map update frequency Improve situational awareness Make faster, data-driven decisions MARS can be run with satellite images from geospatial data platforms, like Microsoft Planetary Computer Pro, providing an end-to-end workflow for automated map production. Users can ingest and manage aerial imagery within Microsoft Planetary Computer Pro, where data is stored and indexed using the STAC (Spatial Temporal Asset Catalog) standard for efficient discovery and search. MARS can perform feature extraction using input data from a user’s Microsoft Planetary Computer Pro GeoCatalog, with the resulting outputs stored back into Microsoft Planetary Computer Pro for downstream analysis. By seamlessly integrating data management and GeoAI inference into a single workflow, MARS and Microsoft Planetary Computer Pro help organizations transform raw satellite imagery into usable maps. Potential Applications Mapping underpins mission-critical workflows across government and commercial industries. When maps become incomplete, outdated, or inaccurate, the decisions and insights built on them can carry significant operational risk. As organizations increasingly invest in digital twins, geospatial analytics, and map-based decision making, there is growing demand for scalable AI-enabled solutions that can transform imagery into GIS map data. Here are several potential applications of MARS. Government and intelligence communities around the globe invest significant resources producing and maintaining maps because roads, buildings, and critical infrastructure are essential to mission planning, reconnaissance, operational logistics, disaster response, and situational awareness. Currently, producing maps often requires analysts to manually review imagery, digitize features, and reconcile information from multiple sources, which is a time-consuming and error prone process that struggles to keep pace with a rapidly changing world. Using MARS, analysts can automatically generate GIS-ready vector outputs from satellite imagery, dramatically reducing the effort required to update maps at a fraction of the cost compared to traditional workflows. Infrastructure operators, utilities, and transportation agencies currently rely on aerial surveys, field inspections, engineering reports, and manually maintained GIS datasets to track changes in the built environment. These approaches can be manual, costly, time-consuming, and difficult to perform frequently across large regions. To address these challenges, MARS provides automated feature extraction from satellite imagery, replacing labor-intensive mapping workflows with scalable, AI-driven map production. This enables organizations to maintain more current geospatial datasets for infrastructure planning, grid modernization, renewable energy expansion, and transportation network management. Telecommunication providers currently combine GIS analysis, third-party mapping datasets, demographic studies, and site surveys to determine where to expand network coverage networks across underserved and rapidly growing areas. However, these datasets can quickly become stale in rapidly growing areas, leading to missed opportunities, inaccurate coverage planning, and costly field validation efforts. MARS enables providers to continuously derive fresh GIS-ready vector data from satellite imagery, helping teams identify service coverage gaps, prioritize infrastructure investments, and accelerate broadband expansion into underserved regions. Mining and natural resource companies frequently use field surveys, drone operations, engineering assessments, and GIS teams to monitor roads, operational infrastructure, and site expansion across vast and remote regions to support mineral exploration, mine planning, and infrastructure management. Updating these datasets often requires significant manual effort and can lag behind changing on-the-ground conditions. MARS can continuously extract and update road networks, waterway features, and operational infrastructure across vast and often remote regions, enabling faster planning cycles and more informed operational decisions. Insurance providers currently rely on stale property records, government datasets, aerial surveys, and manual assessments to maintain property inventories and evaluate risk exposure. MARS can help maintain property inventories and accelerate risk assessment by automatically extracting building footprints and detecting changes from recent satellite imagery. Following natural disasters such as floods, hurricanes, or wildfires, insurers can generate up-to-date maps of buildings, roads, and communities to assess exposure, prioritize inspections, expedite claims triage, and improve portfolio management using current geospatial data rather than infrequent surveys and manually maintained datasets. As organizations increasingly invest in digital twins and smart city initiatives, MARS provides a scalable way to keep virtual representations synchronized with the physical world. Automatically extracting buildings, transportation networks, and other critical assets from new imagery helps ensure digital twins remain current and ready to support planning, simulation, and operational decision-making. Getting Started MARS is available today on Microsoft Foundry. To get started, 1. Deploy the MARS model from the Microsoft Foundry Model Catalog. a. You may need to request additional GPU compute quota in Microsoft Foundry (https://learn.microsoft.com/en-us/azure/foundry/how-to/quota). The MARS model requires the Standard_NC24ads_A100_v4 VM. 2. Set up your Azure Microsoft Planetary Computer Pro instance. 3. Within Microsoft Planetary Computer Pro GeoCatalog, ingest input satellite imagery data as a Spatial Temporal Asset Catalog (STAC) collection of Cloud Optimized GeoTIFFs (COGs). Commercial data is available from various providers and partners like SkyWatch. a. Input Imagery Recommended Specifications: i. Orthorectified satellite or airborne imagery with a ground sample distance (GSD) between 25cm – 75cm (the optimal GSD is 60cm) ii. Ingested into Planetary Computer Pro as an RGB COG and STAC metadata that specifies gsd, projection (e.g. proj:epsg), band information (e.g eo:bands), and accurate bounding box (bbox) and geometry fields b. Recommend Model Settings i. Optimal tiling: 512 x 512 pixels 4. Use the GeoAI SDK to run MARS against imagery from your Microsoft Planetary Computer Pro collection. 5. Generate production-ready vector layers for analysis, mapping, and downstream applications. References AI Model Catalog | Microsoft Foundry Models GeoAI SDK — GeoAI with Microsoft Planetary Computer Pro MARS - A Foundational Map Auto-Regressor | OpenReview Microsoft Planetary Computer Pro | Microsoft Azure Microsoft Planetary Computer Pro | Microsoft Learn
