What a foundation model does to geospatial data
A geospatial foundation model (GFM) learns a compressed representation of spatial context from imagery, sensor feeds and map data. That representation is not neutral: it reflects choices about training data, preprocessing, model architecture, objectives and evaluation, and it carries those choices into every output.
The checks become harder when a GFM brings together many datasets. An output may look convincing while revealing little about the sources, permissions and processing choices behind it. Documentation of training data and transformations is a starting point for trust, not a substitute for it, and relying on a model does not by itself transfer or eliminate responsibility for decisions built on its outputs.
WGIC’s report, Data to Decisions: Rebuilding Trust in Decision-Grade Geospatial Data, proposes an architecture that makes governance part of how data enters the model, how it is used and how the resulting decisions can be inspected.
Three layers, with audit throughout
The report’s proposed architecture connects three layers, with audit spanning the full pipeline.
Layer 1: Data. Inputs can include satellite and aerial imagery, LiDAR, radar, elevation models, vectors, text and mobility data. Their acquisition context, limitations and histories need to remain available as they enter the system.
Layer 2: Governance. A metadata fabric, provenance graph, authenticity gate and sovereignty gate organise information and enforce rules before data reaches the model. Checks address data origins and transformations, licensing, and applicable jurisdictional restrictions. Metadata becomes information the system can query and validate.
Layer 3: Model and decision. A model registry, training and inference processes, outputs and decisions form the final layer. Records connect model use to the relevant versions, input-data classes and controls.
Audit runs across all three layers, supporting inspection from source data through to the decision.
What the checks mean in practice
The report illustrates this through a flood map. Before the map is used to train a GFM, the authenticity and sovereignty gates check its lineage, licensing and jurisdiction. The governance pipeline records its source and model lineage so later outputs can be linked to the relevant model version, input-data classes and applicable controls.
For developers, this means addressing permissions and restrictions early in model development. For users, it creates a basis for examining how an output was produced and whether its use can be justified.
Traceability supports that assessment; it does not establish accuracy or fitness for purpose on its own.
Confidence still needs validation
A model’s confidence score and a geospatial accuracy assessment answer different questions. Accuracy assessment compares results with appropriate reference data. A confidence or uncertainty measure describes uncertainty in a prediction and may itself require calibration and validation.
Governance must address both predictive performance and the reliability of reported uncertainty, particularly when inputs are sparse, noisy or incomplete.
The training data itself also needs protecting. When recursively generated synthetic data contaminates or replaces the original distribution used to train the next generation of models, provenance degrades across training cycles and models can collapse towards their own outputs. Lineage checks at ingestion are the first defence.
Three questions before relying on an output
The report asks users to establish:
- What was the model trained on? Understand how training data was selected, curated and transformed.
- What period does that data represent? Identify whether the model is being applied within familiar conditions or beyond its training period.
- How does it perform for this use? Test against relevant benchmarks and reference data, alongside published leaderboard results.
Building trust into geospatial AI requires checks before data enters the model, records that remain connected through processing, and validation before outputs inform decisions.
Explore the architecture and recommendations in Chapter 5, “Governance by Design”, of Data to Decisions: Rebuilding Trust in Decision-Grade Geospatial Data.
Explore what makes geospatial data fit for a decision. WGIC’s Data to Decisions report identifies provenance, governance, metadata and accountability as the foundations for rebuilding trust, and sets out practical recommendations across the data lifecycle, from data capture and lineage to AI-driven models and decision-making.
It also explores data sovereignty, transparency, uncertainty and the governance of Geospatial Foundation Models.
