Location Is the Join: Why Geography Matters More Than Ever in the Age of AI
After many years of thinking in tables and joins, I've landed on two simple beliefs about geospatial technology. First, location is the most natural join key there is, and that makes it the glue AI needs to fuse very different kinds of data. Second, Tobler's first law of geography quietly underpins a huge share of the analysis that solves real-world problems. In this post, I unpack both ideas using my own work forecasting meningitis outbreaks across Africa.
SCIENTIFIC INSIGHTS INTO AI AND GEOSPATIAL TECHNOLOGIES
11/16/202611 min read


My Two Cents, After All These Years
Tomorrow I'm giving a five-minute lightning talk at GIS Day back home in State College, on forecasting meningitis with predictive geospatial AI.
Five minutes is not a lot of time. So while I was cutting my script down (and cutting it down again, and again), I kept bumping into the same thought:
Most of what makes this work possible isn't the AI.
It's the location.
I've done a lot of work with relational databases, where I spent years thinking about schemas, keys, indexes, and, above all, joins. Left joins, right joins, inner joins, outer joins, and so on. But one of my first loves is geography, where the questions are about place, distance, and how things spread.
If you asked me to boil all those years down to two points about why geospatial matters, here they are:
Location is a natural join. Almost everything happens somewhere, and that "somewhere" lets you connect data that otherwise has nothing in common. As AI gets better at fusing many kinds of data, this matters more, not less.
Tobler's first law of geography is everywhere. "Everything is related to everything else, but near things are more related than distant things." That one sentence sits under an enormous amount of analysis that actually solves real problems.
Neither idea is new. Honestly, that's the point. With all the change happening in the world, and in AI especially, these are two things I'd bet on staying true.
Let me explain why.
Life in Join Land
If you've worked with a relational database, you know the feeling. A question comes in, and before you write a single line of SQL, your brain is already sketching the joins.
Inner join? Left join? Right join? What's the key? What happens to the rows that don't match?
The relational model goes back to Edgar F. Codd's 1970 paper, and the idea is beautifully simple. Store each kind of thing in its own table, and connect tables through shared keys. Customers connect to orders through a customer ID. Orders connect to products through a product ID.
Here's the thing nobody warns you about when you start out: the join is only as good as the key.
If two tables share a clean ID, life is easy. If they don't, you're in for a long week. You match names that are spelled three different ways. You build crosswalk tables. You argue about whether "St. Mary's Hospital" and "Saint Marys Hosp." are the same place. (They usually are. Until they aren't.)
Most data in the world doesn't come with a shared key. A weather model doesn't know your customer IDs. A satellite has never heard of your hospital codes. A news article doesn't come with a foreign key.
But almost all of it has something in common.
It happened somewhere.
Location: The Universal Join Key
This is my first point, and it's the one I've been chewing on the longest.
Location is the one key that almost every dataset already shares.
A clinic report has a district. A dust measurement has a grid cell. A census count has an enumeration area. A conflict event has coordinates. A satellite pixel has a footprint on the ground. None of these datasets were designed to work together. Most of them were made by people who never met and never will.
And yet, because each of them is tied to a place, you can join them.
I call this the Universal Join Key. In SQL, it looks something like this:
SELECT d.district_id, c.cases, p.population
FROM districts d
JOIN conflict_events c ON ST_Intersects(d.geom, c.geom)
JOIN population p ON ST_Intersects(d.geom, p.geom);
The ON clause doesn't compare IDs. It compares geometry. "Does this point fall inside this polygon?" (ST_Intersects in PostGIS is the workhorse here.) Nobody had to agree on a shared ID in advance. The shared key is the Earth itself.
This idea is older than computers. In 1854, John Snow mapped cholera deaths in London's Soho and showed they clustered around one water pump on Broad Street. Think about what he actually did. He took one dataset (deaths, by address) and another (water pumps, by location) and joined them by place. No shared ID. Just a map.
That's a spatial join. Done by hand, with ink, more than a century before Codd wrote his paper.
Why AI Needs a Universal Join Key
So why does this matter more now?
Because AI is heading toward fusion. The most useful AI systems aren't just reading one kind of data. They're combining text, images, sensor readings, tables, and time series into a single picture of what's going on.
And fusion needs a key.
Language models don't solve the join problem. They're great at reading a news report. But to use that report in a forecast, you still need to know where it's about, and line it up with everything else you know about that place. Placing the story on a map is the join.
Earth observation is going "AI-native." In 2025, Google DeepMind released AlphaEarth Foundations, which summarizes satellite, radar, and other observations into a 64-number embedding for every 10-meter square of land, every year. That's a stunning amount of information. But notice what holds it together: each embedding is tied to a location. To use it for anything, you join it to your own data by place.
Grids are becoming shared keys. Systems like H3, a global hexagonal grid, turn location into an actual ID you can put in a column. Two datasets that both use H3 cells can be joined with a plain old equality join. That's location becoming a primary key in the most literal sense.
The pattern is the same every time. The data gets richer and stranger, and location stays the thing that lets it all line up.


The Universal Join Key in Action: Forecasting Meningitis
Let me make this concrete with the work I'll be talking about at GIS Day.
Every dry season, bacterial meningitis sweeps across a band of Africa known as the meningitis belt, from Senegal to Ethiopia. It arrives with the harmattan, the dry, dusty wind off the Sahara. Researchers have linked dust and dryness to meningitis for years. One study found that, during the dustiest months, meningitis peaks were preceded by dust peaks by zero to two weeks.
We're building research forecasts of meningitis risk for about forty African countries: every district, every week, one to four weeks ahead.
To do that, the model needs a picture of each district, each week. And that picture comes from data that has almost nothing in common:
Case reports. Weekly suspected cases, reported by district.
Climate and dust. Gridded data, at a completely different resolution from the districts.
Population. Raster grids of people per cell.
Vaccination. Campaign and coverage information, often at a coarser level.
Conflict. Individual events with point coordinates.
None of these came with a shared ID. But each one has a location and a time. So every one of them gets joined to the same row: this district, this week.
Grids get averaged or summed within each district boundary. Points get counted inside polygons. Coarser data gets passed down to the districts inside it. Different shapes, different resolutions, same key.
Without the universal join key, there is no forecast. There are just five unrelated piles of data.
The Join Tax
Now for the candid part. Location is a natural join, but it isn't free.
Every spatial project pays what I call the Join Tax. Here's what's on the bill:
Boundaries change. Districts split, merge, and get renamed. A district in 2015 may not be the same shape in 2025. If your case data uses one set of boundaries and your map uses another, your "key" quietly breaks.
Unit shape changes your answers. This is the modifiable areal unit problem, described by Stan Openshaw in 1984. Draw the boundaries differently, and you can get different results from the same underlying data.
Coordinate systems don't match. Two datasets can describe the same spot on Earth with different numbers. Join them without checking, and everything is quietly off.
Time is a key too. "This district, this week" only works if every dataset agrees on what "this week" means. Epidemiological weeks, calendar months, and satellite overpass times rarely line up neatly.
And then there's the oldest join question of all: what do you do with the rows that don't match?
In SQL, that's the difference between an inner join and a left join. An inner join drops anything that doesn't match. A left join keeps it, with blanks.
In disease forecasting, that choice is not academic. A district that sent no report this week is not the same as a district with zero cases. If you inner join, you silently drop the quiet districts. If you left join and treat the blanks as zeros, you tell the model nothing happened when you simply don't know. In our work, we flag which zeros are actually missing reports, so the model can tell the difference.
That's a join decision. It just happens to be a public health decision too.
Tobler's First Law: Near Things Are More Related
Now to my second point.
In 1970, geographer Waldo Tobler published a paper about simulating urban growth in Detroit. Tucked into it, almost in passing, was a sentence that became famous: "everything is related to everything else, but near things are more related than distant things."
It's now known as Tobler's first law of geography.
It sounds almost too obvious to be a law. Of course nearby things are related. But once you start looking for it, you see it everywhere:
Weather. Tomorrow's temperature here is a lot like today's temperature a few miles upwind.
Housing. A home's value says a lot about the homes on the same street.
Ecology. Species, soils, and forests change gradually across a landscape, not randomly.
Disease. Infections spread between people who are close to each other, so outbreaks in one place raise the risk next door.
And here's why it matters so much: Tobler's law is the reason so much geospatial analysis works at all.
Interpolation, like kriging, fills in gaps between weather stations because nearby measurements are similar. Spatial statistics like Moran's I measure how much nearby things are related. Hot spot analysis finds places where high values cluster. Spatial regression accounts for the fact that neighbors aren't independent. Even modern graph neural networks, which pass information between connected places, are built on the same intuition.
If near things weren't more related, none of these tools would have anything to grab onto.


Tobler's Law in Action: Your Neighbors Are Your Forecast
Back to meningitis.
When we build a forecast for one district, we don't just look at that district. We also look at what's happening in nearby districts. Proximity to other areas is one of the model's inputs.
Why? Because Tobler was right. If cases are rising in the district next door, that tells you something about your own district's next few weeks. Outbreaks don't respect administrative lines.
This is the neighbor signal. And in practice, it's one of the most useful things a forecast can know.
Our models also lean heavily on recent case reports. When we measure which inputs the forecast relies on most, recent cases come out on top, by far. Climate matters much less once recent cases are known, because recent cases already reflect the season. That's Tobler's law working across time as well as space: near moments are more related than distant ones.
But here's a twist worth sitting with.
"Near" doesn't always mean "close on a map."
A district connected to a big market town by a busy road may be "nearer" in the ways that matter for disease than a district right next door across a river with no bridge. Herders move with their animals across hundreds of kilometers. People travel for markets, festivals, and work.
So the deep version of Tobler's law isn't really about distance in kilometers. It's about connection. Near can mean travel time, road networks, trade routes, or the flow of people between places. One of the most exciting directions for our longer-term work is exactly this: using data on where people move and gather to redefine what "near" means for disease spread.
Tobler's law doesn't break. It just gets a better ruler.
Honest About the Gaps
I don't want to oversell any of this.
Our forecasts are a research prototype. They're not deployed, and the early results are under peer review. Those early results are encouraging: compared with existing methods, our forecasts are better calibrated, meaning that when the model says a district has a 30% risk, outbreaks happen about 30% of the time.
But we're also candid about what we can't do yet. When we looked back at past epidemics, most weren't flagged as high-risk one to four weeks before they began, by any method we tested, with the data and models we've tried so far. Districts can go from quiet to epidemic in a week or two, and case reports often arrive late.
Here's the part that connects back to this post. When we ask what would close that gap, the answer keeps coming back to the same two ideas:
More data, joined by location. Better vaccination records. News reports as early signals. Data on movement and gatherings. Every new source is only useful if we can join it to the same district and week.
Better ways to use nearness. Models that capture how disease actually travels between places, not just between neighbors on a map.
The universal join key and the neighbor signal aren't just how we built the forecasts. They're the roadmap for improving them.
Why This Matters Beyond One Disease
Step back from meningitis for a second.
Climate change is shifting where and when diseases show up. Extreme heat, floods, and droughts don't hit everyone equally; they hit specific places. Migration, displacement and conflict reshape where people live. Supply chains, food security and energy are all, at their core, questions about place.
At the same time, AI is changing how we work with data faster than I've seen in my whole career.
When everything is changing at once, it helps to know what isn't.
Here's what I'd bet on:
Data will keep getting more diverse. New sensors, new satellites, new kinds of text and images. Location will remain the key that joins them.
The world will keep being spatially connected. Problems will keep spreading through neighbors, networks, and flows. Tobler's law will keep describing them.
That's why I think geospatial skills are more valuable now, not less. If you can join messy data by place, and if you understand how nearness shapes what happens, you can bring AI to bear on problems that matter.
And to any students reading this, especially the ones I'll see tomorrow: the people who will build the next generation of these tools are people like you.
Conclusion: The Map Is the Join
After all these years, I keep coming back to two ideas:
Location is a natural join. It lets us combine data that was never meant to be combined, making it the backbone of AI that fuses diverse data.
Near things are more related. Tobler's first law sits underneath so much of the analysis that actually solves real-world problems, from weather to housing to disease.
John Snow joined deaths to pumps with a map. Today we join case reports, dust, population, vaccination, and conflict data by district and week. Tomorrow we'll be joining embeddings from foundation models we haven't even imagined yet.
The tools change. The key doesn't.
What's Next?
Thanks for reading! If you're in State College, come say hi at GIS Day. And if you're not, you can watch online.
Now I'm curious: what's the hardest spatial join you've ever had to do? And where have you seen "near" mean something other than distance on a map?
Let's keep the conversation going on LinkedIn!
Recommended further reading:
Waldo Tobler (1970), A Computer Movie Simulating Urban Growth in the Detroit Region, Economic Geography: the paper where the first law first appeared.
Edgar F. Codd (1970), A Relational Model of Data for Large Shared Data Banks, Communications of the ACM: where joins began.
London Museum, John Snow: Cholera & the Broad Street pump: the original spatial join.
Martiny & Chiapello (2013), Assessments for the impact of mineral dust on the meningitis incidence in West Africa, Atmospheric Environment.
