Flai spent Intergeo 2026 in Munich, September 15–17, at our booth in Hall C6, E106. Three days of live classification demos, a steady stream of visitors, and a lot of conversations with people we've gotten to know over several editions of the show — mapping company PMs, national agency contacts, hardware and software partners. Intergeo is one of the few places the whole LiDAR and geospatial community shows up in one hall at the same time, and this year had that feel: busy aisles, repeat faces, and the kind of conversations that pick up where last year's left off rather than starting from scratch.
A lot of what came up at the booth was familiar, which is its own kind of signal. Teams are still fighting the same fight — data volumes growing faster than headcount, QC time eating into delivery schedules, pressure to automate the repetitive 80% of classification without losing control over the 20% that needs a human eye. What's changed is how matter-of-fact those conversations have gotten. Automated classification isn't a hard sell anymore; the questions are about deployment, accuracy at scale, and fitting it into an existing pipeline.
That shift from "should we automate" to "how do we automate at our scale" was also the throughline of Flai's session in the Smart City track: "AI-Powered LiDAR Classification at National Scale: Inside the Digital Twin Germany Programme," given by Flai's Blaž V. It used one of the largest coordinated LiDAR campaigns in Europe — Germany's Digital Twin (DigiZ-DE) programme — as a concrete answer to that question.
The scale problem: 356,794 km² at 40 points per m²
DigiZ-DE is led by Germany's Federal Agency for Cartography and Geodesy (BKG). The programme is surveying the entire country — 356,794 km² — at a point density of 40 points per square meter, roughly four to eight times denser than a typical national LiDAR survey. The output is a unified 3D model of Germany, built to support urban planning, crisis management, and long-term sustainability work.
Higher density means better detail. It also means the data volume scales faster than the team classifying it can. A workflow that holds up on a single county-sized project doesn't automatically hold up when you multiply it by a country, and a QC process built around manual review can't keep pace with acquisition that runs continuously across hundreds of thousands of square kilometers. At this scale, automated classification isn't a convenience — it's the only way the throughput math works.
An independent overview of the programme, published in GIM International, is available here.
How classification works across a national programme
Flai's role in DigiZ-DE is AI-powered point cloud classification, applied in a single automated run across ground, vegetation, buildings, water, bridges, power lines, wind turbines, and additional project-specific object classes.
A few things the talk covered that are specific to running classification at this scale, not just at this volume:
- Strip-level noise classification. Noise doesn't behave the same way across a country-sized dataset acquired over months and multiple flight campaigns. Classifying noise at the strip level, rather than applying one blanket filter to the whole dataset, keeps quality consistent as acquisition conditions vary.
- Granular power line categorization. Power infrastructure needs finer categories than a single "power line" class once you're covering a full national grid — separating wire types and structures rather than lumping them together.
- Project-specific object classes. Wind turbines are a good example: not a standard ASPRS category, but a class BKG needed classified consistently across the programme. Extending the base scheme to cover cases like this is part of making automation work for a specific client's requirements, not just a generic dataset.
- A 12-class ASPRS-based scheme, extended with custom classes. Rather than building a bespoke schema from scratch, DigiZ-DE builds on the standard 12-class ASPRS scheme and adds the categories the programme actually needs. That keeps output compatible with existing GIS and survey tooling while still capturing the detail a national digital twin requires.
Local vs. cloud processing, at national scale
One trade-off the talk addressed directly: whether to run classification locally or in the cloud. For a programme moving petabyte-scale data across hundreds of thousands of square kilometers, the answer isn't the same as it is for a single regional project. Bandwidth, data residency, and turnaround time all shift the calculation differently depending on where the data is captured and who needs to see it first. There's no universal answer — it depends on the acquisition cadence, the infrastructure already in place, and what the delivery deadlines actually require.
Why this matters beyond DigiZ-DE
DigiZ-DE is one programme, but the pattern isn't unique to Germany. National mapping agencies and regional surveying authorities across Europe are moving toward higher-density, more frequent LiDAR coverage — and running into the same bottleneck: classification throughput that doesn't scale with acquisition throughput.
The session was aimed squarely at that audience — national mapping agencies, regional surveying authorities, and geospatial production teams running large LiDAR programmes — because the questions DigiZ-DE raises (how to classify consistently across strips and campaigns, how to extend a standard schema without breaking compatibility, where to run the compute) are the same ones any large-scale programme eventually has to answer.
What's next
DigiZ-DE is an active, ongoing programme — classification work continues alongside acquisition as BKG's survey coverage expands. Flai's role in the programme will keep growing with it.
For us, Intergeo 2026 was a good measure of where the market's head is at: less time spent convincing people automation works, more time spent working through the specifics of how to run it at their scale. Thanks to everyone who stopped by the booth or sat in on the session — see you at the next one.
If your organization is running, or planning, a large-scale LiDAR campaign and thinking through the same classification and deployment questions — including on-premise or air-gapped requirements common in government and defense contexts — get in touch with the Flai team.
Frequently Asked Questions
What is the Digital Twin Germany (DigiZ-DE) programme?
DigiZ-DE is a national LiDAR survey programme led by Germany's Federal Agency for Cartography and Geodesy (BKG). It's mapping all 356,794 km² of the country at 40 points per square meter — four to eight times denser than a typical national LiDAR survey — to build a unified 3D model of Germany for urban planning, crisis management, and sustainability applications.
How does AI-powered LiDAR classification scale to national or country-level point cloud volumes?
The main requirements are consistency and automation depth: classifying noise, ground, and object classes automatically at the strip or campaign level rather than relying on manual review, and extending a standard classification scheme with project-specific classes so output stays usable across the whole programme without bespoke rework each time acquisition conditions change.
What classification scheme works best for national-scale LiDAR programmes?
DigiZ-DE uses a 12-class ASPRS-based scheme extended with project-specific classes, including categories like wind turbines that fall outside the standard ASPRS set. Starting from an established standard and extending it, rather than building a fully custom schema, keeps the output compatible with existing GIS and survey tooling.
Should large LiDAR programmes process data locally or in the cloud?
It depends on the programme. Bandwidth, data residency requirements, acquisition cadence, and delivery deadlines all factor into whether local (on-premise) or cloud processing makes more sense, and large national programmes often need both, depending on the dataset and the client's infrastructure. There's no single right answer independent of those constraints.
How is power line classification handled at national scale?
Rather than a single generic "power line" class, national-scale programmes typically need power infrastructure separated into more granular categories — distinguishing wire types and structures — so the output is usable for grid management and inspection, not just visualization. See our post on automated power line classification for how this works in practice.
Where can I see Flai's classification accuracy benchmarks?
Flai publishes independent accuracy evaluations against real production data and hardware, including F1 scores by class and processing time. See our evaluation of Flai AI classification on data from the AISPECO Heliux LITE system and RIEGL VQ-580 II-S for a worked example.