What is spatial intelligence?

A MODIS cloud climatology published in IEEE Transactions on Geoscience and Remote Sensing puts the mean global cloud fraction at about 67%, and about 55% over land. So a five-day optical revisit can deliver one clean scene a month. The lidar campaign may be five years old, and the BIM model shows the building as designed. Every source is partial, and each one goes out of date at a different rate. Spatial intelligence is the engineering work of fusing them into one model that records how old each part is.

A working definition of spatial intelligence

The term is used loosely, so it helps to pin it down. Spatial intelligence is not a data source and not a piece of software. It is a capability: the ability to answer questions about a physical place, with a known level of accuracy, at a known point in time. "Which drainage channels in this district silted up since March?" and "Has the footprint of this storage yard changed since the last survey?" are spatial intelligence questions. Answering them needs data, alignment, interpretation and history, all in one place.

Four properties separate a spatial intelligence system from a folder of maps. First, it is georeferenced: every pixel, point and polygon carries coordinates in a defined coordinate reference system (CRS), so layers line up. Second, it is fused: imagery, point clouds, design models and sensor readings describe the same objects rather than sitting side by side. Third, it is queryable: users and software can ask spatial and attribute questions through an interface instead of opening files. Fourth, it is time-aware: the model knows when each observation was made and can show what a place looked like at any date.

Time awareness is the property most often missing. A traditional map is a snapshot that silently ages. A spatial intelligence model records the acquisition time of each input, the time each feature was last confirmed and the history of changes. That lets a user ask not only "what is here" but "how sure are we that it is still here" and "when did it change". For operational teams, the second and third questions usually matter more than the first.

The data layers and what each adds

Each data source has a characteristic trade-off between spatial resolution, revisit frequency, coverage and cost. Satellites cover large areas often but at coarse detail. Drones and aircraft capture centimetre-level detail but only when someone flies them. Lidar measures geometry directly. CAD and BIM describe design intent rather than observed reality. IoT sensors report conditions continuously but at single points. A spatial intelligence model uses each source for what it does well and lets the others fill the gaps.

The table below summarises the main layers. Satellite and lidar figures are taken from the operators' own documentation.

The Sentinel-2 figures come from the Copernicus Sentinel-2 mission description, which lists four 10 m bands (B2, B3, B4, B8), six 20 m bands and three 60 m bands across a 290 km swath. The Landsat figures come from the USGS Landsat 9 mission page, which notes that Landsat 8 and 9 are offset by eight days. The radar figures come from the Copernicus Sentinel-1 mission description. Lidar point density and vertical accuracy follow the USGS 3DEP topographic data quality levels.

Two points from the table deserve emphasis. Revisit time is not the same as usable revisit time: a five-day optical revisit over a cloudy region may yield one clear image a month. And resolution is not accuracy. A 10 m pixel can be positioned very precisely, while a 2 cm drone orthomosaic without ground control can be offset by metres. A good model tracks both properties separately for every layer.

From raw data to a queryable model

Turning raw inputs into a usable model follows a consistent chain. The order matters, because errors introduced early propagate through every later step. A misaligned image produces false change detections, and a misclassified point cloud produces a wrong terrain model. The steps below describe the chain as most mature geospatial teams run it.

1. Ingest: collect imagery, point clouds, design files and sensor streams, and record their metadata (sensor, acquisition time, processing level, licence).
2. Georeferencing and CRS alignment: place every input in a common coordinate reference system.
3. Fusion: link observations from different sources to the same real-world objects.
4. Feature extraction: detect and classify objects and surfaces, usually with machine learning.
5. Change over time: compare states and record what changed, when and with what confidence.
6. Serving: expose the model through APIs and open standards.

Ingest is less trivial than it sounds. Satellite data arrives at different processing levels: for Sentinel-2, Level-1C is top-of-atmosphere reflectance and Level-2A is atmospherically corrected surface reflectance. Mixing the two produces inconsistent indices. Free Sentinel archives are available through the Copernicus Data Space Ecosystem, but the ingest layer still has to handle tiling, cloud masks and metadata so that later steps know exactly what they are working with.

Georeferencing and CRS alignment is where most quality problems start. Global data often arrives in WGS 84 (EPSG:4326) with geographic coordinates, while engineering work needs a projected CRS in metres, such as the relevant UTM zone or a national grid. Drone imagery needs ground control points or RTK/PPK GNSS positions to be tied to the ground. Optical satellite imagery needs orthorectification against a DEM so that terrain relief does not shift features. CAD and BIM models are often drawn in a local site coordinate system and must be transformed using surveyed control points. Until every layer agrees within a known tolerance, fusion is unreliable.

Fusion then links observations. A building footprint from imagery, its roof height from lidar, its room layout from BIM and its energy meter from an IoT feed should resolve to one object with one identifier. In practice this means spatial joins with tolerances, rules for which source wins when values disagree, and a record of where each attribute came from. Lidar processing is a sub-chain of its own: points are classified into ground, vegetation and buildings, a DTM is interpolated from ground returns, a DSM from first returns, and subtracting one from the other gives a normalised height model for canopy or structures.

Feature extraction and change detection

Raw pixels and points are not answers. Feature extraction turns them into objects and classes: roads, roofs, water bodies, tree crowns, crop fields, solar panels. For many years this relied on spectral indices and rule sets. NDVI, computed from red and near-infrared reflectance, separates vegetation from bare surfaces. NDWI and MNDWI, which use green with near-infrared or shortwave infrared bands, highlight open water. These indices remain useful because they are cheap, interpretable and physically grounded.

Most production feature extraction now uses machine learning. Random forest classifiers are still common for pixel-based land cover on multispectral stacks, because they handle many bands, train quickly and give feature importances. For object boundaries, convolutional networks dominate. The U-Net architecture described by Ronneberger, Fischer and Brox was designed for biomedical segmentation, but its encoder-decoder structure with skip connections became a standard for semantic segmentation of building footprints, roads and water in aerial and satellite imagery. Point clouds use their own network families, such as PointNet++ and KPConv, for per-point classification.

Change detection compares two or more states of the same place. Simple approaches difference indices or classifications between dates. Learned approaches train a network on image pairs. Daudt, Le Saux and Boulch's fully convolutional Siamese networks process both dates through shared weights and predict a change map directly, and the authors frame the work as a step towards processing data from large-scale Earth observation systems such as Copernicus and Landsat. SAR adds another route: because Sentinel-1 sees through cloud, backscatter changes can flag flooding or clearing when optical sensors are blind.

Model quality has to be measured, not assumed. Segmentation is usually reported with intersection over union (IoU) per class, and detection and change tasks with precision, recall and F1. Change detection is heavily imbalanced, because most of a scene does not change between dates, so overall accuracy can look excellent while the model misses most real changes. Per-class recall on the change class is the more honest figure. Any spatial intelligence system should store these metrics with the outputs so that users can judge how far to trust a given layer.

How spatial intelligence differs from traditional GIS

Geographic information systems are the foundation spatial intelligence builds on, not a competitor to it. GIS provides the data models (vector and raster), the coordinate handling, the spatial operations and decades of cartographic practice. The difference lies in how the data is maintained and consumed. Traditional GIS workflows are largely project-based: an analyst collects layers, processes them on a desktop, produces maps or reports, and the work ends.

Spatial intelligence treats the model as a living system rather than a project output. New imagery arrives on a schedule, pipelines process it automatically, change is detected and recorded, and the model updates without an analyst rebuilding it. Machine learning replaces much of the manual digitising that GIS teams used to do. The audience also changes: alongside GIS specialists, the users are operations managers, field crews, engineers and software systems that query the model directly.

The interface shifts accordingly. Desktop GIS centres on files such as shapefiles and GeoTIFFs. Spatial intelligence centres on services. The OGC API - Features standard is a good example: it lets clients discover, query and retrieve individual features over HTTP rather than downloading whole datasets, and Part 1 is also published as ISO 19168-1:2020. Map tiles, coverages and catalogues follow similar web patterns. This is what makes a model usable by dashboards, mobile apps and machines, not only by people who know GIS software.

How it differs from a static twin

"Digital twin" has become a broad label. In construction and facility management it often means a detailed 3D model of an asset, built from BIM and handed over at completion. Standards such as ISO 19650 for information management using BIM govern how that information is exchanged and versioned across the asset's life. A model like this is extremely valuable, but on its own it describes how the asset was designed and built, not how it is today.

A static twin degrades from the day it is delivered. Equipment moves, temporary structures appear, vegetation grows, extensions are built without a model update. After a few years, the twin and reality can disagree in ways that matter for maintenance, safety and planning. The model still looks precise, which is part of the risk: users trust a clean 3D view that no longer matches the site.

Spatial intelligence turns a static twin into an operational one, a distinction we cover in more depth in what is a digital twin. It keeps the BIM or CAD model as the structural backbone, then layers observed data on top: drone surveys, lidar scans, satellite change detection and live sensor feeds. Each new observation either confirms the model or flags a difference. The result is a twin with a known currency for every part of it. Where a static twin answers "what did we build", an operational twin answers "what is there now, and when did we last check".

Who uses spatial intelligence

Municipalities are among the largest users. City and regional governments need current maps of buildings, roads, drainage, green space and informal construction for planning, taxation and service delivery. Change detection on satellite or aerial imagery can flag new structures for review instead of relying on periodic manual surveys. Lidar-derived terrain models feed flood risk analysis, and linked attribute data lets departments share one base map rather than maintaining separate copies.

Utilities use spatial intelligence to manage networks spread across large areas. Electricity operators monitor vegetation encroachment along power line corridors using lidar canopy heights and satellite vegetation indices. Water utilities map reservoirs, canals and treatment sites, and track water extent with NDWI and SAR. Solar park operators use drone thermal and RGB surveys to locate faulty panels. In each case the value lies in combining network records with observed conditions.

Forestry and conservation teams use multispectral time series to map forest cover, detect clearing and monitor regrowth, with SAR filling cloud gaps in humid regions. Lidar gives canopy height and structure that optical sensors cannot measure directly. Agriculture uses similar tools at field scale: NDVI time series indicate crop vigour and stage, field boundaries come from segmentation models, and soil moisture or weather sensors provide ground truth that calibrates the satellite signal.

Ports and industrial sites have a different mix of needs. Their areas are smaller, but the density of assets and the pace of change are high. Container yards, stockpiles, quays and warehouses change daily. Here drone surveys, fixed cameras, CAD layouts and asset tracking matter more than wide-area satellite coverage. Spatial intelligence gives operations teams a single view of where things are, how much material sits in each stockpile and whether the layout still matches the plan. You can see more examples on our use cases page.

Limits and failure modes

Cloud cover is the most familiar limit: nominal and usable optical revisit can differ by weeks. SAR helps because it is unaffected by cloud and darkness, but it measures surface roughness and moisture rather than colour, so it cannot replace optical data for every task. Speckle noise, layover in steep terrain and geometric distortion also need careful processing.

Resolution sets a hard floor on what can be detected. A 10 m Sentinel-2 pixel cannot reliably show a narrow footpath, a single small vehicle or a crack in a roof. Mixed pixels at object boundaries blur area estimates. Higher resolution sources exist, but they cost more, cover less and are captured less often. Choosing the right resolution for the question, rather than the highest available, is a core design decision in any spatial intelligence programme.

Ground truth is the limit that teams most often underestimate. Every classification and change model needs labelled examples from the area it will be used in, and validation data collected independently of the training set. Models trained in one climate or building style often degrade elsewhere. Field visits, survey data and local sensor readings are slow and expensive to collect, but without them the accuracy figures reported for a model mean very little.

Data currency is the final limit, and it is easy to hide. A model fused from a five-year-old lidar campaign, last month's drone survey and yesterday's satellite image has three different ages in three different places. If the interface does not show this, users will assume everything is current. A well-built system attaches acquisition dates and confidence to every feature, and treats "unknown since" as information rather than as a gap to paper over.

What it means for robots

Robots build their own world models with SLAM (simultaneous localisation and mapping), fusing lidar, camera and inertial data into an occupancy grid or point cloud map. That works within sensor range and within one mission. It works less well when several machines share a site, when the site changes between missions, or when a machine arrives somewhere for the first time and has no map at all. A shared, georeferenced model fills those gaps in three concrete ways.

The first is the prior map. An autonomous mobile robot (AMR) in a warehouse can be given a 2D occupancy grid generated from the CAD layout and the latest scan of the racking, then localise by matching its lidar returns against that grid instead of exploring first. Outdoors, a ground robot uses a DTM and a lidar-derived point cloud as the prior. For this to work, the robot's local map frame must be tied to the model's projected CRS through a known transform, so a position in metres on site is also a position in UTM or a national grid.

The second is mission planning. A drone survey needs a DSM to set flight height above buildings and canopy, a DTM for terrain-following, and polygons for restricted zones. A ground robot needs traversability, derived from slope, surface class and mapped obstacles. A survey boat needs the shoreline as a geofence and bathymetry to keep clear of shallows. All of these are layers of the same spatial model, read through the same API.

The third is change triggers. When satellite change detection flags a new structure or a cleared area, the flagged polygon can become the next drone mission, and the affected map tiles can be marked as stale for every robot that relies on them. When robots then report what they see, the model updates and the loop closes. Our robotics solutions page sets out how we see this layer developing.

How Kaiinos approaches spatial intelligence

Kaiinos builds spatial intelligence for people and autonomous machines. The team has more than a decade of geospatial delivery work, including municipal, forest, water, agriculture and utility mapping programmes. That work shaped a practical view of where projects succeed and fail: alignment and data currency cause more problems than model architecture, and outputs only get used when they reach operations teams in a form they can query.

Our Smartout platform is the operational digital twin layer that puts this into practice today. It processes CAD/BIM, lidar and drone data, integrates satellite imagery, fuses multiple sensors into one georeferenced model, and provides visualisation, spatial queries, change history and APIs. It is used to create and maintain digital views of environments such as warehouses, manufacturing sites, municipalities, forest areas, solar parks, fisheries and maritime sites. The technology page explains the processing stack in more detail.

The next layer we are building towards is a robot perception engine. The aim is to generate robot-ready twins from CAD/BIM, sync them to robots, consume their camera and lidar feeds, detect local changes and update the shared twin in near real time. Beyond that, we are working towards an edge module that runs on robots, drones and ground vehicles, and longer term towards navigation intelligence, mission planning and fleet coordination. These are roadmap directions, not current features: navigation and autonomous driving decisions are not part of Smartout today.

Frequently asked questions

What is spatial intelligence in simple terms?

Spatial intelligence is the ability to answer questions about a physical place accurately and at a known point in time. It combines satellite, drone, lidar, design and sensor data into one aligned model, uses machine learning to identify objects and changes, and makes the result available through software interfaces so people and machines can query it directly.

Is spatial intelligence the same as GIS?

No. GIS supplies the data models, coordinate handling and spatial operations that spatial intelligence depends on. The difference is in operation: traditional GIS work is usually project-based and analyst-driven, while spatial intelligence runs as a continuously updated system, with automated pipelines, machine learning feature extraction, change history and API access for users beyond GIS specialists.

Which satellite data is used most for spatial intelligence?

Free Copernicus and Landsat data form the base for many programmes. Sentinel-2 provides 10 m optical bands with a five-day revisit at the Equator, Landsat 8 and 9 provide 30 m data with a long archive, and Sentinel-1 radar sees through cloud. Commercial high-resolution imagery is added where finer detail is needed.

How accurate is a spatial intelligence model?

Accuracy depends on each input and on processing quality, so it should be reported per layer. Airborne lidar at USGS quality level 2 targets 10 cm vertical RMSE in non-vegetated areas, while satellite-derived features are limited by pixel size. Machine learning outputs should carry IoU, precision and recall figures validated against independent ground truth.

How does spatial intelligence help robots?

It gives robots a shared, georeferenced world model instead of isolated local maps. A current prior map supports localisation, mission planning and coordination between machines on the same site. When robots feed their own observations back, the shared model can stay current. At Kaiinos this robot-facing layer is on our roadmap rather than shipped today.

Build a site model that stays current

If you are planning a mapping, monitoring or digital twin programme, the hardest questions are usually about alignment, data currency and validation rather than software. We are happy to go through your data sources and requirements in detail. Book a technical call with our team.