What if communities could map their own worlds using low-cost drones and open AI models instead of waiting for expensive satellite imagery?

In this episode with Leen from HOT (Humanitarian OpenStreetMap Team), we explore how they’re putting open mapping tools directly into communities’ hands—from $500 drones that fly in parallel to create high-resolution imagery across massive areas, to predictive models that speed up feature extraction without replacing human judgment.

Key topics:

• Why local knowledge beats perfect accuracy
• The drone tasking system: how multiple pilots map 80+ square kilometers simultaneously
• AI-assisted mapping with humans in the loop at every step
• Localizing AI models so they actually understand what buildings in Chad or Papua New Guinea look like
• The platform approach: plugging in models for trees, roads, rooftop material, waste detection, whatever communities need
• The tension between speed and OpenStreetMap’s principles
• Why mapping is ultimately a power game—and who decides what’s on the map

In Conversation

What HOT Does and Why Community Involvement Matters

Daniel: You’re the director of data and technology at HOT. Can you tell us what HOT does and give us a project example?

Leen: HOT — the Humanitarian OpenStreetMap Team — is focused on improving livelihoods through open map data. We build open-source technology to enable that. In practice, we engage communities in creating and using map data for their own challenges — flooding, wildfires, disasters, waste mapping. We work with communities to map together, not just map for them. When people are part of generating the data, they also take ownership of using it. The goal is to give citizens agency over the maps that represent their world.

Daniel: Couldn’t you get much of this data from satellites? Wouldn’t that be more accurate?

Leen: Accuracy isn’t always the most important element — it depends on the use case. If there’s a disaster, you just need to know which buildings are destroyed. You don’t need sub-centimeter precision for that. And satellite imagery only captures what sensors can see. It doesn’t tell you if elderly people live in a building, what type of doctors are available at a hospital, or what the local community actually needs. Those details only come from the community itself. Imagery is critical for mapping, but local knowledge is what gives it meaning.

The Imagery Gap and Drone Technology as the Answer

Leen: One of the key gaps in our work is that there’s no recent, high-resolution imagery accessible in many of the areas where we operate. If we want to run a project in Ghana or Freetown and we need current imagery at a useful resolution, it’s often unavailable, or available only at very high cost from providers like Maxar and Airbus. We do receive post-disaster imagery from Maxar’s open data program, but only after a disaster has already occurred. We want to work pre-disaster too — on development projects unrelated to catastrophe — and that’s where the gap is.

Leen: Drone technology is becoming a commodity. Prices are dropping and the hardware is increasingly accessible. We’ve been investing in it for several years now. Low-cost, lightweight drones can be used for mapping by the communities themselves. We’ve even been experimenting with building our own hardware using 3D printers — very experimental, very low-cost. But the bigger initiative is our drone tasking manager, which enables communities to fly multiple drones in parallel, each covering a different area, then stitch the results together into high-resolution imagery across a much larger area than any single drone could cover.

Parallel Flight: 80 Square Kilometers in Freetown

Daniel: Can you walk me through what that looks like in practice?

Leen: Last month we finalized this approach for the first time at scale — 80 square kilometers in Freetown, Sierra Leone. A project coordinator creates the flight plans. Each pilot takes a section, flies in parallel with others, and uploads imagery when their drone returns. All of that imagery gets processed using open-source tools — primarily WebODM and OpenDroneMap — and then pushed to OpenAerialMap, where anyone can view and download it. The whole workflow uses open tools and produces open data. We’re planning to replicate this in Dar es Salaam, Dhaka, and Malawi over the next six to eight months.

Daniel: Right now this still requires a computer. What’s the plan to simplify it further?

Leen: Today there’s still quite a bit of friction — USB cables, laptops, manual uploads. In a year, the goal is to make the entire workflow mobile-only. A drone pilot with just a phone and a drone should be able to receive their flight plan, fly it, upload imagery directly from their phone, and have it pushed to OpenAerialMap with a single button — all processing happening in the background. No computer required. That matters enormously in the countries where we work, where reliable connectivity and access to laptops can’t be assumed.

AI-Assisted Mapping with Humans in the Loop

Daniel: Now that we have imagery — how do you turn it into map features? How does AI-assisted mapping work?

Leen: Today, a lot of manual mapping happens where people draw polygons, points, and lines on imagery. We’ve been optimizing this with a pre-processing step using an app called MapSwipe, where people swipe left or right on their phone to indicate whether there’s something worth mapping in a tile. Then in our Tasking Manager, people trace buildings and other features, with contributions going back to OpenStreetMap wherever possible.

Leen: What we’re now doing — and have proven results for — is using open AI models to extract features automatically, with humans validating and correcting the results at every step. Open here means open code, open license, open training data. The AI assists; the human decides. People train the models, generate training data, and validate predictions. It’s much faster, especially post-disaster, but it doesn’t replace human judgment — it augments it.

Localizing Models for Different Contexts

Daniel: If you use a general AI model, does it understand what buildings look like in Chad or Papua New Guinea?

Leen: No — that’s the critical problem. General models are typically trained on data from western countries. They have no idea what a round hut in Chad looks like, or the building density and styles of informal settlements in South Asia. So we take an open-source general model and fine-tune it with specific local training data — outlines of buildings in Chad, for example — to create a localized model that actually works in that area. Community members generate the training data by drawing polygons. A project manager with some technical knowledge handles the model parameters and accuracy metrics. Then the community validates the predictions. Anyone can participate — in principle, someone could swipe on their phone to say yes or no on a predicted building without any technical background.

Opening the Platform to Any Geospatial AI Model

Leen: The next step is turning our open AI mapping tool into a platform that anyone can plug their model into, rather than us implementing models ourselves. Using the new OGC ML Model standards, a model developer — a university, an NGO, anyone — can contribute their open AI model for detecting waste, road surface types, seagrass, fish ponds, or rooftop materials. Project managers can browse available models, select the one that fits their use case, localize it with training data, and run predictions — all through the same interface, without needing engineering expertise.

Daniel: Does HOT curate the models, or can anyone plug one in?

Leen: We’re starting by working with universities already developing open-source geospatial AI models. Over time, the vision is that the community itself rates and recommends models based on what worked in their context — if a model worked well in an area similar to yours, you should be able to reuse it. The goal is to lower the barrier for mappers and project managers so they don’t need to be AI engineers to use the right tool for their community’s needs.

The OSM Tension and Who Controls the Map

Daniel: How does all of this fit with OpenStreetMap? Is there tension there?

Leen: There is some tension. OSM’s principles are built around human-driven community mapping, and AI-assisted mapping creates friction — both in principle and in practice. Some community members question whether speed is really the goal, or worry that AI predictions aren’t good enough. And honestly, today’s AI isn’t always better than a careful human mapper. But we believe the process is still fundamentally human: humans generate training data, humans validate predictions, humans push results to OSM. The pushback often comes from mappers in the western hemisphere. In the countries where we work, communities just want the data — they want to be on the map, regardless of whether the building outlines are perfectly squared.

Leen: And whether it ends up in OSM or not, this data is incredibly valuable. We publish on platforms like HDX, the UN OCHA open data portal, where organizations download country data when responding to crises. Being on the map matters. It’s a power game — who decides what gets mapped, where, and how. That’s what we’re trying to shift: communities deciding for themselves.