Satellite change detection for municipalities

At 0.5 m ground sample distance, a 4 m by 3 m rear extension covers about 48 pixels. On imagery from Sentinel-2, the free Copernicus satellites whose sharpest bands have 10 m pixels, it is about an eighth of one pixel. That difference decides whether satellite imagery can keep a municipal building register accurate. A register changes only when someone files a permit or an inspector passes. The city changes anyway: extensions go up, front gardens are paved, solar panels appear and trees come down. Very high resolution imagery records these changes whether or not anyone filed the paperwork.

Records that no longer match the ground

A municipal spatial data infrastructure combines base maps, building and address registers, parcel boundaries and the systems that read from them, such as permits and property tax. Each dataset has its own owner and update cycle, and gaps open up between them and what is built: a building exists on the ground but not in the register, or a recorded footprint was enlarged long ago.

These gaps have direct costs. Records that miss extensions understate the tax base, and enforcement teams see unpermitted construction only after a complaint. Climate adaptation suffers too: when gardens are paved and trees are cut down, rainwater has nowhere to soak in. A European Commission staff working document on soil sealing states that sealing increases the vulnerability of areas to climate change, resulting in increased flood risk.

Manual surveys cannot keep pace: by the time inspectors finish the last district, the first is out of date again. The task is to find the places that have probably changed, so field time goes there first.

What we built across three city corporations

Kaiinos delivered a municipal project across three city corporations over three years. The spatial data infrastructure was incomplete and out of date, and buildings on the ground did not match the records. The aim was to bring the records back in line and keep them there.

We built three connected parts. The first was building identification from satellite imagery, which produced building footprints that could be compared with the existing records. The second was a single spatial database that holds the building data and serves it to other municipal systems through APIs, so permits, tax and planning teams work from the same layer instead of keeping their own copies. The third was an offline-capable mobile app for field teams. Inspectors receive flagged locations, verify them on site, and send their findings back to the database.

The third part matters as much as the first: a model output is a suggestion until a field team confirms it. Our Smartout platform is built around the same ideas, with change history, spatial queries and APIs over operational digital twins of places such as municipalities. It is one application of the broader spatial intelligence approach. In Europe, we are applying it to municipal satellite imagery, comparing images over time with each municipality's own base maps and building and address registers.

Choosing imagery: resolution, revisit and cloud

The imagery source is a trade-off between ground sample distance (GSD), revisit, cost and weather. According to the Sentinel-2 mission documentation, its finest bands (blue, green, red and near-infrared) are 10 m, the red-edge and SWIR bands are 20 m, and two satellites give a 5-day revisit at the equator. A 10 m pixel covers 100 m², so a typical house occupies only a handful of pixels.

For building detection, very high resolution (VHR) optical imagery at sub-metre GSD is the working standard, because it resolves the extensions and outbuildings that registers miss. Sentinel-2 still has a role at district scale, for example tracking vegetation loss through NDVI or broad land-use change.

Cloud is the other constraint: in cloudy climates, clear optical acquisitions can be scarce for weeks. Synthetic aperture radar does not need daylight or clear skies; ESA describes Sentinel-1 as providing an all-weather, day-and-night supply of imagery. SAR is better used as a trigger for closer inspection than as a source of building outlines.

Preprocessing: making two images comparable

Change detection only works if two images line up pixel for pixel and show the same surface in the same way. Three steps get them there. Orthorectification removes terrain and sensor-geometry distortion, using a digital elevation model and the sensor's orbit and attitude data, and places the image in a defined coordinate reference system. Without it, a building on a slope or near the image edge appears shifted, and the shift looks like change.

Co-registration aligns the two dates to each other, usually to sub-pixel accuracy, by matching tie points or by cross-correlation. A misalignment of one or two pixels at 0.5 m GSD is enough to produce false edges along every building in the scene. Radiometric normalisation then corrects differences in illumination, atmosphere and sensor response, so that a roof that has not changed has similar values on both dates.

Seasonal effects remain: leaf cover, sun elevation and shadow length differ between June and December. Comparing same-season images, or training on pairs that include these differences, reduces false change.

Building footprint extraction

Building extraction is a segmentation problem. Semantic segmentation assigns each pixel a class, such as building or background. The U-Net architecture, introduced by Ronneberger, Fischer and Brox in 2015 for biomedical images, became a common baseline in remote sensing because its encoder-decoder structure with skip connections keeps sharp boundaries. Instance segmentation goes a step further and separates individual buildings. Mask R-CNN does this by adding a mask branch to an object detector.

Instance output fits registers better, because registers store buildings as separate objects, while semantic masks of terraced housing merge into one block. Either way, the raster is vectorised and simplified into polygons, often with corner-squaring rules, before comparison with existing footprints.

The SpaceNet dataset and challenge series released labelled satellite imagery on AWS, and its first two competitions focused on automated building footprint extraction. Models are usually scored on intersection over union (IoU), with a predicted footprint counted as correct above a set IoU threshold, and then summarised as precision, recall and F1.

Change detection: two families of methods

There are two broad ways to detect change. Post-classification comparison classifies each date separately, for example by extracting building footprints from both images, and then compares the two maps. A 2020 review of AI-based change detection by Shi and colleagues notes that this structure copes well with images from different acquisition conditions and even different sensors, but its accuracy depends on both classifiers. Errors on either date carry through to the change map.

Direct bi-temporal methods feed both images into a single model that predicts change directly. Siamese networks are the common design: two sub-networks with the same structure extract features from each date, and the features are compared by concatenation or a distance measure. The same review distinguishes pure Siamese networks, which share weights, from pseudo-Siamese variants, which do not. Direct methods are usually more sensitive to small changes but need training pairs that contain real change.

The LEVIR-CD dataset by Chen and Shi is the standard benchmark for building change. It contains 637 bi-temporal image pairs of 1024 by 1024 pixels at 0.5 m, with more than 31,000 labelled building change instances and time spans of 5 to 14 years per pair. In practice, municipal systems often combine both families: a direct model finds candidate change, and footprint extraction describes what the new or altered building looks like.

Reconciling detections with registers

A change map is not yet useful to a municipality. It has to be compared with the records. The core operation is a spatial join between detected footprints and register footprints, followed by matching rules. A detected building with no register footprint nearby is a candidate new building. A detected footprint that overlaps a register footprint but is clearly larger is a candidate extension. A register footprint with no detection is a possible demolition, or a building hidden by trees.

The same logic applies beyond buildings. New paving in a garden parcel, solar panels on a roof or a lost tree crown are matched against parcels, the building register or a tree inventory. Thresholds on overlap, area change and model confidence decide which matches count, and they need tuning per city, because building styles and register quality differ.

Prioritisation ranks the matches by size of change, model confidence and relevance. A large unrecorded structure matters to permits and tax; a paved garden in a flood-prone street matters to climate adaptation. Each department sees its own queue from the same layer.

The field verification loop

Field verification is where detections become records. In the project, field teams used an offline-capable mobile app, because mobile coverage cannot be assumed everywhere an inspector goes. In this kind of setup, each flagged location arrives with the imagery evidence, the existing record and the detected change. The inspector confirms, rejects or corrects it on site, and the result syncs back to the spatial database when a connection is available.

This loop does three jobs. It prevents action on false positives, because no record changes without a human check. It produces confirmed and rejected examples, which make the best training data because they reflect the municipality's own buildings and imagery. It also makes performance measurable: the share of flags confirmed in the field is a direct estimate of precision, and shows whether thresholds are set well.

Limits and data protection

Resolution sets a floor on what can be seen. Small changes such as a garden shed, a dormer window or a narrow paved path may fall below what sub-metre imagery reliably resolves. Viewing angle is another limit. The SpaceNet MVOI study used 27 views of the same area from -32.5 to 54.0 degrees off-nadir and found that leading segmentation and detection models struggle with off-nadir imagery and generalise poorly to unseen angles. Tall buildings lean, and their facades cover neighbouring ground.

Shadows, tree canopy and cloud also hide change. A flagged extension may already have a permit, or the register footprint may have been drawn badly. Output should go to field teams and case officers, never straight into a tax assessment or enforcement notice.

For EU municipalities, data protection applies from the start. Building data linked to addresses and owners can relate to identifiable people, and the General Data Protection Regulation defines personal data in Article 4(1) as any information relating to an identified or identifiable natural person. Article 35(3)(c) requires a data protection impact assessment for systematic monitoring of a publicly accessible area on a large scale. A municipality should define the legal basis, purpose and retention period, and complete that assessment before linking detections to owners.

From change layer to robots

A verified change layer is a set of dated polygons, each with a change type and a confidence score. Autonomous machines can use that format directly. Drones are the most direct consumer: each flagged polygon can become a survey area: the flight planner buffers it, sets altitude from a DSM so the drone clears nearby roofs, and captures oblique images or lidar. Those resolve what satellite imagery cannot, such as building height, a new storey or a dormer. The satellite layer decides where to fly, and the drone delivers the detail. A district-wide flight becomes a short list of targeted missions.

Ground robots and vehicles use the same layer as a map invalidation signal. A sidewalk delivery robot or road maintenance vehicle localises against a prior map of kerbs, walls and building edges. Where that geometry has changed, scan matching degrades and the robot's position estimate drifts. A change layer, queried by the fleet's operating area through an API, tells the map team which tiles to resurvey and tells the fleet where to slow down or reroute. Simulation teams need it too, because a synthetic test city is only realistic if its base model is current. See how this connects to our robotics work and autonomous systems.

Today, Smartout maintains operational digital twins with change history and APIs that other systems can query. The next layer we are building towards is a robot perception engine that consumes camera and lidar feeds from machines in the field, detects local changes and updates the shared twin in near real time. Satellite change detection would then be one input among several, with robots reporting what they see back into the same model.

Frequently asked questions

Can Sentinel-2 imagery detect new buildings?

Sentinel-2's finest bands are 10 m, so a single house covers only a few pixels. It shows broad patterns such as new development areas, vegetation loss or land-use change, but cannot reliably outline individual buildings or extensions. For building-level work, municipalities use very high resolution optical imagery at sub-metre ground sample distance.

How often should a municipality run change detection?

It depends on how fast the city changes and how much field capacity exists. A practical starting point is annual or twice-yearly acquisition, in the same season to reduce false change. Running more often than field teams can verify adds little, because unverified detections do not update any records.

Does change detection replace building inspectors?

No. It changes where inspectors spend their time. Instead of surveying every street, they visit locations that the system has flagged as probably changed. Their confirmations and rejections update the register and also improve the model, so the two work together rather than one replacing the other.

Is satellite change detection compliant with GDPR?

It can be, if designed for it. Building data linked to addresses and owners can be personal data. A municipality needs a clear legal basis and purpose, limited retention and a data protection impact assessment before linking detections to individuals. A human verification step before any decision also reduces risk.

Start with one district

The fastest test is a pilot on one district, using your own base maps and building register with two dates of imagery. You see what the system flags, what your field team confirms and what it would take to scale. Explore other use cases or book a technical call to scope a pilot.