The New Science of Maps

What it is, and why I built it

I have been receiving many questions about how to start with geospatial data science, which learning materials to use, and which tutorials to use, from intro to advanced GeoAI topics. Any my reply usually was a selection of links, some of which got outdated, or simply lost among the many platforms, from YouTube to Medium, from Amazon to Substack. Sometimes even I got lost in all the different channels I used to share educational and research materials.

And this was my clear go-sign: I need a unified platform collecting the geospatial data science work I did during the past years. This is how New Science of Maps was born - the platform where all my courses and books now live, but it is also a claim about what this field has turned into. So here is the longer answer on what it is — this time without any coding.

The hard part is no longer the code

Looking back at those scattered links, what strikes me is that the problem itself is quite new. Ten years ago there would not have been much to collect. The hard part of spatial work back then was access: GIS software cost money, satellite imagery cost more, and the skills mostly lived inside a few university departments and consultancies. Nobody was drowning in tutorials, because there were hardly any.

All three of these barriers have essentially disappeared since. Sentinel-2 is free and comes back every five days. GeoPandas, OSMnx, rasterio, H3, PyTorch — all free, all mature, all well documented. And now a language model will write your reprojection code faster than you can remember whether you need to_crs or set_crs, which one of the two silently mislabels your data, and why that difference will quietly ruin every area calculation downstream.

That last one is where it gets interesting. The tools have become remarkably good at doing things, while they haven't really become any better at telling you whether the result means anything. So the scarce skill has moved somewhere else: from running the analysis to the judgment around it. Which method fits this question, what a dataset can and cannot support, and whether a perfectly plausible-looking result is in fact wrong.

An example from my own work. A few years ago I looked at green equality in Vienna — how much accessible green space people have per capita across the city's 250 census districts, and whether that relates to the income level of the district. I expected a fairly clear pattern here, since that is broadly what the literature reports elsewhere. Instead I got almost nothing: a Spearman correlation of 0.13, and a Pearson of 0.30 that I trusted even less, given how heavily skewed the income data is.

Every line of that code ran without a single error. The analysis itself was fine. What actually mattered was everything around it — choosing which correlation to take seriously given the shape of the data, and then deciding that "no relationship" was a real finding about a city with decades of deliberate greening policy, rather than a failed analysis. Honestly, I was a bit surprised by the result, and that surprise turned out to be the most useful part of the whole project.

This is the difference I keep seeing between an analysis that survives contact with a client and one that falls apart in the first meeting. It is almost never a coding problem. It is whether anyone stopped to ask if the number could be true.

Why I call it a New Science of Maps

A map used to be a representation. You surveyed something that existed, and you drew it.

What we do today is mostly inference. Take a livability map of a city, colored by neighborhood — something I have built many times. It is not a picture of anything that exists out there. It is the visible end of a long chain of decisions: which indicators you include, how you weight them, which spatial unit you use and whether that unit does any analytical work or just happens to be where the administrative boundaries fell, how you normalize, and what you do with the areas that have no data at all. Change any single one of these, and the map changes. And every version of it was produced honestly.

This is not cartography with better software. It is modeling, in the sense data scientists or even Physicists would call it, with a map as the output format. It has its own failure modes, its own limits on what you are allowed to claim, and its own ways of being confidently wrong in front of an audience.

I think it deserves to be taught as its own thing, from first principles, by someone who will also tell you where it breaks. That is what the name is claiming — not a new invention, but a way of seeing that quietly became a discipline while nobody gave it a name.

What is actually on the platform

Courses and books on established geospatial data science journeys: urban analytics, GeoAI, satellite data, network science, and the Python foundations underneath all of it.

A few things I care about, and that make this different from a pile of video lessons:

Real study areas, real data. Vienna, Manhattan, Edinburgh, rural Hungary, the Netherlands. Actual OSM extracts and actual Sentinel-2 scenes, with the messiness left in — not a cleaned teaching dataset that behaves nicely.

From first principles. Here I am going back to my roots as a Physicist and aim to learn the foundations and how to build things from scratch, rather than accepting black boxes. This is not the most efficient way to work day-to-day, but the safest in the long run, especially with AI helping do work for us on every corner. But once you have done it once, every abstraction built on top of it stops being magic, and you can tell when one of them is lying to you.

Make it quick. While I start from the basics, I also acknowledge that today, no one - including me - can afford to spend months on a single topic. Hence the structure of my books: 101 steps, which are optimized for efficiency; and the length of my courses, measuring just a few hours from start to finish.

Built to be taken in an order. The courses are designed as paths that ladder into each other, with the Python foundations sitting underneath so the rest do not have to keep re-teaching them. I am putting these paths together properly on the site right now.

Two versions of every notebook. One commented version you will actually use, and the exact notebook from the recording — so when I type something on screen, you can find that literal cell and see what it did.

Named limitations. Where a method stops working, I say so during the lesson, not in a footnote afterwards.

And what it is not

It is not a tour of software. There are excellent people teaching specific tools, and I am not trying to compete with them.

It is also not complete, on purpose. There is no SQL or PostGIS here, no survey of cloud-native formats, and no full treatment of SAR. All three are genuinely important for a lot of working practitioners — they are simply not what this platform teaches, and I would rather say that clearly than pad the catalog and let you find out after buying something. However, this field is still evolving rapidly - and also relies on your feedback, so some of these may change in the future

What is coming

More courses, and the learning paths finished properly. A cohort-based certification twice a year is under planning. Live sessions where I work through real problems in real time, including the parts that do not go well on the first attempt. And eventually a single way in for people who would rather have everything than pick.

I will write about each of these once they are actually ready, not before.

Where to start

Everything lives at thenewscienceofmaps.com. There are free samples and a runnable companion notebook there, and the fastest way to find out whether any of this is for you is to open one and run it.

If you want the rest as it comes — tutorials, new courses, and the research I am working on — this newsletter is where it all goes first. Thanks for reading, and see you in the next one!

— Milan