Posts on Dani Arribas-Bel

What Battersea Power Station taught me about the value of satellite embeddings

Tue, 05 May 2026 09:59:51 +0100

Sometimes, magic finds its way through the forest of math that underpins most of my research. I spent a good amount of time a couple of weeks ago working on our embeddings workshop for GISRUK. We just released a new data product with Google satellite embeddings for small areas in the UK, and we couldn’t wait to take it on the road. This post is a note about one of the most insightful moments I had while preparing the materials but, if you want to check out the entire workshop, it’s all open at:

https://imago-sdruk.github.io/embeddings_workshop/jupyterlite/content/02-Change.html

One of the most appealing characteristics of satellite embeddings is their ability to explore change. Since embeddings map the information in an image to a shared latent space, it doesn’t matter where or when that image was taken, it all gets encoded in the same mathematical “language”. Exploring whether the location of an image is similar or not to that of a different image; or whether such location has changed compared to what it looked like at a different point in time, is relatively straightforward. You can play with the “comparison across space” idea in our Imago UK embedding Explorer. And we played with change across time in this notebook.

The part that got me is the last exercise. We ask the question what area has become the most like Hyde Park between 2020 and 2024? This seems a bit esoteric but, I think, is an interesting one to ask. We know areas change within a city all the time, wouldn’t it be cool to know which ones are becoming more like some specific landmarks? In the workshop, I called it “Hyde Park-ification”, but I may deny this in public… It also turns out this is relatively straightforward to check with embeddings. You can follow all the details, code included, in the notebook. The gist is you calculate how similar all areas are to Hyde Park in both years, take the difference of that distance, and pick your winner as the smallest of those.

Here’s our winner¹:

For the uninitiated reader, this is Battersea Power Station. Which, even if all you know is the first photo in the Wikipedia page I just linked to, you’d be forgiven to think this embeddings sorcery is nonsense. This is as far as you can get from Hyde Park in terms of look and feel. So, where’s the catch?

The catch, of course, is that we are not looking into overall similarity, but change. Our approach looks for areas that, over the period we consider, have become more like Hyde Park. That does not mean they are like Hyde Park. There’s another interesting bit we show in the notebook: Battersea Station, by the standards of the areas we use, has changed very little. Again, this is counterintuitive maybe, but not incompatible with what we’re asking of embeddings in this exercise: you can change very little, but entirely in one particular direction.

As it turns out, this is what happened to our winner. This insight did not come painlessly to me, as I struggled through a good 20/30 minutes of scratching my head about whether I was asking too much of this technology (particularly when it’s aggregated from the pixel to an irregular area). I’m glad I stuck with this and followed it to the bottom. And I’m glad Google keeps funding Google Earth for free, including its time travel feature. Below is a comparison of the area in question between 2020 and 2024²:

If you can’t spot the difference (I couldn’t for the first ten minutes), follow the red arrow (which wasn’t there when I was looking for twenty minutes). The area, broadly speaking, has not changed much (which checks out with our low score of change). It was, and still is, a retired power station turned epicenter of cool south of the river. Mostly chimneys, mostly concrete. But there’s a small part that has changed. A small patch by the river used to be concrete buildings but, in between 2020 and 2024, it was flattened. Instead, it is now a patch of grass where Londoners enjoy the three days a year where the sun shines with all its fury³. You can see some of the before and after photos in this article by the BBC.

I love this example because it captures very well why I’m so excited about embeddings and about making them accessible to more people through data products like our small area one. I knocked out that notebook in about an hour (if you exclude my Google Earth rabbit hole). In that period, I was able to leverage data from seven massive satellite sources (see Figure S2 in this paper for details), across two periods of time on a setup that runs on my iPad’s browser. Just let that sink in. What questions can we tackle that we couldn’t before? More importantly, which ones will we actually consider now that the price of asking is so low? I think we’re about to find out, and I can’t wait.

I’m pasting here screenshots because it’s 2026 and YOLO, but you can play with the interactive map on the workshop site. ↩︎
You can play with the feature here. ↩︎
This is a joke. I know it’s, at least, five days. ↩︎

Notes from a week in the Bay Area

Mon, 23 Mar 2026 13:01:00 +0100

I’m just wrapping up a week in Berkeley and San Francisco. I came for two reasons, both have been fantastic, and I’m definitely feeling now the post-high blues that good academic interaction tends to induce.

I spent the first few days at the Berkeley Institute for Data Science, where the good folks of PySAL organised the Spatial Data Science Summit. This was a meeting to discuss opportunities and overlaps across the spatial data science community, and beyond, including the wider Python ecosystem for scientific computing. These days, I’m much less active (and that’s a kind way of putting it) in the development of PySAL. But it was nevertheless a fantastic experience. PySAL 1.0 shipped in 2009 and I was there to see it. Since then, it’s been a project and a community that has brought me much of what I am and who I’ve become. And, as the cliché goes, many of the collaborators have become close friends. Clichés are clichés for a reason.

Then I hopped on the BART to San Francisco’s Tenderloin, where geographers were descending left and right to discuss all things Geography at the AAG. I was part of a session brilliantly put together by Elizabeth Delmelle and Geoff Boeing on “problem-driven methods” (as opposed to the seemingly more common “method-driven problem development”) in the context of cities. I presented on satellite embeddings. When I was preparing the slides, I felt a bit uncomfortable because, arguably, I was about to engage in precisely the type of behaviour the sessions sought to avoid: “reaching for this year’s shiny new tool” instead of addressing cities' ‘wicked problems’”. Elizabeth and Geoff prompted us to include a slide at the beginning stating the “big urban problem” we were tackling. I’m so glad they did because it really nudged me to spell out why I think embeddings are more than this year’s shiny tool. My framing revolved around two key arguments. First, there is much more we can (and should!) do to tailor the data we use to the problems we tackle. Second, there is a lot of untapped data to help in that tailoring. In this context, satellites are one of those underutilised sources that can help provide better empirical fit to the questions we care about. And embeddings lower the barrier to access satellite data, making it cheaper to ask questions and explore ideas. I think, in the end, it went well and was well received.

I also participated in a panel organised as part of the discussion we had started in Berkeley earlier in the week around open source in spatial data science research. Serge Rey prompted us to think about existing gaps, low-hanging fruit, and surfaces of overlap. We covered quite a bit of territory, and Serge structured the conversation so there was clear and constant interaction with the audience, which became an “additional panelist”. A lot of fun.

Besides the strictly “work” things, this week also had plenty of space for fun. We (i.e., Rachel, Levi and your truly) attended the AAG Awards ceremony to pick up our Media Achievement award for GLaD. To celebrate it, that afternoon we hijacked a boardroom, stuck a hand-written note in the door that read “recording in process”, and taped the first episode in a long while where we were physically together. And, after all that flurry of activity, the day after, I managed to convince Geoff Boeing and Martin Fleischmann to join me on a walk around very tall trees north of the Golden Gate. Long walks are underappreciated ways to exchange ideas.

With that, the week came to an end. It’s been so much compressed in such a small amount of time and space. I know it’ll take me a few days to unpack, literally and figuratively. This is conferencing at its best: more brain cycles in less days, all away than your usual routine. Academic life the way it should be.

2025 in reads

Mon, 29 Dec 2025 14:18:14 +0100

It’s that time of the year (again). Last year was the second time and, if I manage a third one this year, we can start calling it a tradition. And I quite like traditions like this. Before we get into the good stuff, one clarification. The sharp eye will have noticed I did drop “book” in this year’s title. This is all but unintended. In part, it is a bit of ego caressing¹, in part it might be something a little more meaningful. Earlier in the year, I signed up for Reader and, later on, I picked up an Onyx Boox. All of a sudden, the whole of the internet was presented to me like a book I could take to bed without distractions. That’s taken its toll on actual books², but it’s not been for nothing. I have read more long-reads and discovered more perspectives than I’d have been able by focusing only on books. So, this year, I’m giving them a spot on my review by highlighting the ones that have made a mark as of December (way too many to list wholly, but you can see a sample in my link blog). With that out of the way, enough of intros, links and footnotes (you can see a full list of books and links at the bottom), let’s do it.

This year, there’s not been a clear winner on non-fiction, like last year. There’s been a few really good ones that I’ve returned to in a less obsessive way over the subsequent months. I greatly enjoyed What if we get it right? and it might be one of the few instances where I think the audiobook is probably a better medium. It’s set up as interviews with a bunch of people from all corners of life on what “getting it right” with the climate would look like. The audiobook contains the actual interviews, not a readout of them. Not all of them touched me in the same way, but enough of them made me think in slightly different ways about climate change. I also really liked How infrastructure works. It’s less of a thrilling read, more of a slow, boring in the best possible sense of the word (calm, undistracted) one. If you stick to it and get past some aspects of the tone, there’s plenty to like and learn on the other side. One of them was the insight of how the challenge in infrastructure for the XXIst Century is going from a world where resources felt unlimited but energy limited to one of “infinite energy and finite materials”. And the core message of the book is also one to retain with you: that infrastructure is, at its best, shared care, humanity coming together to care for each other at scale. Try to find that elsewhere.

It was also a good year of fiction. No five stars but plenty of fours. I got around (pun intended!) to read Orbital and thought it was very well written but, perhaps, not more. I liked Delta-V because it came at a time where I was really in the mood for good, “hard” sci-fi, it’s pretty much that and it delivers at it. And speaking of traditions, I returned to the Slough House series for my yearly fix, ahead of the TV version. All in all, most very satisfactory, almost none much of a surprise. The closest I came to it was Bluebird, Bluebird, which was a new author to me and a new detective series set in contemporary East Texas. It was a nice mix of good detective execution and a great context setting in terms of race, culture and idiosyncracies of that corner of the world.

And now for the ”new” medium, the Internet. According to Reader, I’ve highlighted (my measure of whether a text had anything I wanted to remember or not) 262 items this year. Many of them were technical, irrelevant by now (AI!) or not that memorable. A few of them have managed to stick to my brain in ways that have surprised me. About half of them were about AI in one way or another, there was sooo much of it this year, some of it was even good! I read a lot of Cory Doctorow, so it’s no surprise something from him made the list. I’m a bit saturated of “enshitification” at this point, but his new big idea, Reverse Centaurs and how AI is being designed to be helped by humans rather than the other way around is as interesting as it gets, I imagine we’ll hear more of it in 2026. I finally got around reading (listening to!) AI as Normal Technology and can confirm all the chatter around it is justified. I enjoyed particularly the first of the three parts, which focuses on the economics of the technology and how it will (or not) spread throughout the economy. This is econ 101 but, sometimes, you “just” need the basics well understood. Speaking of understanding well, Neal Stephenson has some thoughts too (and a new newsletter!). From his post on AI I took the “every augmentation is also an amputation” quote (which was actually from Marshall McLuhan), and it’s returned to me every time I read a new shiny app feature description that promises to make me so much more productive. Harper Reed is one of the few folks occupying the shared space in a Venn Diagram with people excited about AI and people I like and respect, and his “codegen” hero’s journey is equal parts insightful and fun(ny). A bit of a technical one, so indulge me but, if you are into satellites, the TerraWatch newsletter is a treat in your inbox. His recent take on why privatising basic Earth Observation is a terrible idea is a good entry point. To wrap up, two very different and very personal reads that involve absolutely zero AI or technical stuff. I don’t know how I came across Mike Monteiro, but his newsletter has stuck in my read digest. I imagine it’s the combination of honesty, rawness and the occasional insight. His recent post about flying to bury his father (after a very complicated history) is all those three condensed and multiplied by several x’s. And the final one, a very apt one at that, is by Craig Mod. I’ve read a lot of him (no less his phenomenal book on walking in Japan) this year, so it’s also no surprise he makes the list. This one is a retrospective on the house in Tokyo he’s selling. But it’s really a meditation about all the time he’s passed there, what’s meant to him, and how the house was both context and sometimes even content to much of it. It’s beautiful.

And that does it for this year! If you(‘ve) read any of the above, remember next time you see me one of the few things I like more than reading is talking about reading. If there are others I should read, hit me up with recommendations. Life’s too short to fill it with mediocre reads.

Fiction

Delta-V by Daniel Suarez
Rose/House by Arkady Martine
La asombrosa tienda de la señora Yeom by Kim Ho-yeon
Foundry by Eliot Peper
Simplicity, by Mattie Lubchansky
London Rules by Mick Herron
Orbital by Samantha Harvey
Bluebird, Bluebird by Attica Locke
Ice’s End by P. Finian Reilly

Non-Fiction (* for audiobooks)

What If We Get It Right?* by Ayana Elizabeth Johnson
How Infrastructure Works by Deb Chachra
A City on Mars* by Kelly and Zach Weinersmith
The Anxious Generation* by Jonathan Haidt
Techno rebels - The renegades of electronic funk by Dan Sicko
China in ten words by Yu Hua (with translation from Allan H. Barr)
Mood Machine* by Liz Pelly
The Illegals by Shaun Walker
Things Become Other Things by Craig Mod

Internet reads

The Reverse-Centaur, by Cory Doctorow (and my post on it)
AI as Normal Technology, Arvind Narayanan and Sayash Kapoor (and my post on it)
Remarks on AI from NZ, by Neal Stephenson
An LLM Codegen Hero’s Journey, by Harper Reed
Why “Science-as-a-Service” doesn’t work for Earth Science by Aravind Ravichandran (and my post on it)
Studio Goodbye, Studio Hello, by Craig Mod
How to bury your father, by Mike Monteiro

According to Goodreads, I have, after all, read less books and less pages than last year. To be exact, 5,751 instead of 6,671, and 18 books over last year’s 19. I need to tell myself this is for a Reason! ↩︎
For the record, I do think there’s value in actual books. The general criticism of “kids these days” is that they take too many pages to deliver the message. That’s the whole point! They force you to make space to think about an idea in ways you can’t with almost any other medium. ↩︎

Dispatch from the Human Planet Forum

Mon, 24 Nov 2025 12:28:58 +0100

I’m just back from my first Human Planet Forum. In its own words:

The Human Planet Forum 2025 is a flagship event of the GEO Human Planet Initiative (GHP) which works to understand and map human presence on earth using open geospatial data.

In my own, it’s an (the?) opportunity to see in one room, at the same time, all the wonderful nerds behind some of the most interesting data products using earth observation (EO) to map the world’s population. As an understatement, I was excited.

The organisers kindly asked me to prepare a ten minute summary of the event that’d help conclude it. As my email inbox will attest, I “had to” pay attention to all presentations, and thus the opportunity cost of getting distracted was higher than usual. I’m glad I was given this constraint because it made me appreciate the event and what was presented there much more. Below is a write up of that summary. Heads up, this is not a detailed account of each presentation (for that, you’ll have to wait for the report the JRC will put out in time), more of an opinionated view of what the event represents and why it is valuable.

I structured my contribution along three key dimensions: domains discussed, perspectives used, and common threads. I then added two extra slides, one with themes that will keep this community busy for the foreseeable future, and one with my own perspectives to wrap up.

Start with the domains. There was an important chunk of time spent discussing new ways and products that characterise environments. A lot of this is the evolution from the earlier work defining the extent of human settlements (the HS in GHSL!), and it’s very welcome to see these attempts becoming more sophisticated, adding more depth and providing more detail than what it has already been shown possible. It was not surprising to see a lot of work also towards identifying, measuring and monitoring disasters, vulnerability and exposure. Underlying it all was the looming presence of climate change and the consequences it is increasingly having on human systems. And one of the key distinctions between this meeting and almost any other one I can think of is its ambition to combine these domains with socio-economic aspects. Of course, population estimates is by now a staple of this community, but we also heard from efforts to characterise the building stock, the degree of urbanisation, or poverty, to name a few.

I identified three key perspectives. One is presentations on advances in the field, technical or otherwise. What are the things we can now do we couldn’t a few years ago? The second is a focus on introducing foundational layers. The word foundation has lost much meaning in the past few years “thanks” to the AI hype around LLMs¹, but its use here referred to something more meaningful: data layers built to develop a lot of applications on top of them, to be used by a wide range of stakeholders in a variety of contexts, and relied upon for many years to come. This aspect of temporal continuity for me deserves special praise. While the Valley mentality of “disrupting”, using the latest and greatest to put out catchy data products has its (mostly demonstrational) value, to me it pales in comparison with the commitment to support a data product over time.

All of the above stems from a series of shared values, a common narrative, that permeated through pretty much everything that was said in the meeting. The first is the very clear ambition to develop and use data to inform real decisions. There was a very interesting debate across different sessions of how much this ambition should dictate the data we develop and how we go about this task. The appeal to tailor something as close as possible to what decision makers state they need is clear. Less clear to me until I heard it was the notion that maybe such stakeholders are not always aware of what is possible and hence, when they specify needs, miss opportunities. It is then up to us to grab such opportunities to show new ways in which data can influence policy. There was also a lot of discussion about the need for capacity building, for addressing that “last-mile delivery” challenge to make sure the data and insights we develop have real impact. I half joked that, in some cases, I’m not sure how long that mile is. Sometimes, it feels our outputs should inform decisions, but in reality there’s a broader gap between them and the evidence needed to make a difference. At any rate, that this conversation featured so prominently across sessions seems to me like like a very healthy pre-condition to figure out the best way forward. And, finally, the third common thread was the value of Open. Everyone here understands that, for this work to be relevant, it is imperative it be openly available through permissive licenses. This is in contrast to other academic fora where my sense is that openness is seen as a desirable though not critical feature of work. I think the Human Planet folks are “on the right side of history” here.

After touching on domains, perspectives and common threads, I reflected on themes that, in my view, will remain relevant in the coming years, despite all progress already made. The first was the tension between developing global products and aiming to influence local decisions. This is what I called the everything, everywhere, all the time goal of the data products we develop. This is not only an aspirational goal, it is one whose goalposts are constantly shifting (what was “high resolution” or “frequent” five years ago is no longer now), so the need to always push the envelope does not go away with technological advances. To me the real value of the goal is not achieving it per-se (it is always going to somehow elude us), but to help focus our minds about the direction of travel and the “itinerary”: what are the pieces we really cannot sacrifice and which ones we can give up in the name of actually existing datasets? The second point I raised was that notion of “last-mile” discussed above. Figuring this one out will keep us entertained for a while. And the third was the need to continue standardising and harmonising data and approaches to make the ecosystem more useful than the sum of its (data product) parts. As more datasets are created, this becomes more important, but also more onerous.

I closed with three thoughts this Human Planet Forum has stimulated. First is this notion I’ve been mulling over for a while of EO as a data source of first resort. For a long time, at least in social science, one would use satellites because the data one was really interested in wasn’t available (a census, a survey, a building cadaster….). I think the time is ripe for this to change. Data from satellites (and the techniques we now have to work with them) have advanced to a point where I think, in many cases, they make for the best option. This is not fully obvious to everyone (yet) and perhaps it needs a bit more of shouting from people like us to spread this gospel. The second is the value, not only of pushing the boundaries of what is possible with these data, but also to bring such boundaries to the masses. So much of science gets “locked” in niche communities because it is too complicated, obscure, or there is no effort to disseminate. We cannot afford this with advances in satellite data. In this context, I really take away the idea of the “last-mile” as a neat way to encapsulate much of this challenge (and, perhaps, part of its solution). The final reflection I made is that there’s a lot of opportunity in bringing in conversation work along these lines of the Global North with that of the Global South. A lot of the focus in this community has historically focused in less developed countries (see the point about data of last resort above). As technology gets better and becomes more relevant not as a substitute for another data, but as the main focus, I think there’s a great opportunity for all the lessons learned while working on the Global South to be shared and amplified elsewhere. Yet more avenues for cross-pollination and collaboration.

And that was it. As I mentioned at the start, these were three very intense but extremely productive days. Michele and team did a stellar job in putting a fantastic show together and I’d like to thank them again for letting me be part of it. I hope to see many of the participants again very soon in different venues!

Ironically, an area I expected to hear more here was foundation models for satellite imagery but, alas, it was barely mentioned. Perhaps for the better, although I’d have very much enjoyed seeing this community’s take on the new wave of computer vision models trained on satellites for satellites. ↩︎

(Satellite imagery) embeddings for the rest of us

Fri, 24 Oct 2025 14:52:00 +0100

Earlier this week, I attended a workshop on Geo Foundation Models put together by Weiming Huang and Nick Malleson, both at Leeds. It was a small group, high interaction day with plenty of coffee breaks to follow up con comments, and discussions as close to blue sky as academia permits these days. As the cool kids say, extremely high signal to noise ratio.

My contribution was a five minute talk on why embeddings (and, in particular, embeddings from satellite imagery) are a very cool technology we should be paying more attention in social research and policy. I titled it, in an attempt to sound clickbait’y, “(Imagery) embeddings for the rest of us”. The point I wanted to make was that, in my view, embeddings are one of the coolest (new) ways we have to democratise access to much of the value that satellite imagery has to offer. This is particularly so for communities who have not been able to engage much with but stand to benefit from satellites. But I’m getting ahead of myself. In this post, I wanted to give a quick overview of what those five minutes were. Here you go.

I started framing embeddings from Imago’s perspective. At Imago, we work to make satellite imagery more useful, useable, and used across social research and policy. A big part of this is about developing data products that translate pixels into data that meet our users “where they are”. That is, we take relevant information from pixels and provide it in familiar formats (e.g., Census geographies), in transformed/aggregated ways (e.g., tabular) that resonate more with how social mindsets would think.

Then I moved on to embeddings. It was a bit silly to include a slide on what embeddings are for a room full of experts on this area. Nevertheless, I did it because I thought it’d be useful to frame how I see embeddings in this context. As such, I defined embeddings as the internal representation a neural net builds from an image. This ends up being a vector of values that provides a dense but compressed representation of the statistical information encoded in an image. In more human-friendly terms, this is a bit like “an image, as seen by a computer”.

Now, here’s where the talk starts getting more interesting (hopefully). Why are embeddings, such an obscure property of modern neural architectures, so important for socially-minded folks? In my view, at least for three reasons. First, it’s a very direct translation of what is essentially a multi-dimensional tensor (an image) into tabular format. The embedding is a flat, one-dimensional (mostly) representation of an image¹, a very complex data structure. It’s a “buffer” that helps the non-initiated user to not have to touch a raw image, with all of its challenges, but to still get most of the benefits of doing so. Think of it as having the cake and eating it. Second, embeddings are not limited to a single type of image. Modern models are multi-sensor (i.e., they can incorporate different satellite feeds) and even multi-modal (e.g., combining satellite with other data types such as traditional geographic features). To me, this is an opportunity to make many data feeds, traditionally left aside for being “too hard to work with”, a first-class citizen in “tabular-land”, and maybe even a way to integrate seamlessly different communities that’d usually not speak to each other. And third, it’s 2025, we’re talking about embeddings because they’re tightly linked to foundation models, pre-trained general purpose encoders that can generate such embeddings easily once available. This is a really exciting field to observe these days. In the talk, I included a slide with examples of such models released (mostly) openly from Google, NASA, ESA, IBM and Cambridge. All of those examples were, at most, three months old.

As an illustration, I gave a few examples of what one can easily do with these embeddings. They’re ideal for “semantic search”, where you are interested in finding locations that look, in fundamental ways, similar to another one (e.g., what is another area in the UK that looks like my neighbourhood in Liverpool?). They’re natural inputs for unsupervised classifications built with K-means or more modern algorithms (e.g., what are the key types of areas in this region?). And they have a lot of potential to help easily spot change (e.g., has this area changed between the two periods for which we have images?). Of course, these are the “standard” uses embeddings are being used for. One of the things I’m really excited about taking embeddings to a much broader audience of domain experts is seeing what they can do with them to help solve their specific challenges.

I closed the talk reflecting a bit why this is not commonplace yet. In particular, I brough up three thoughts. The first one is that what I had just said sounds likely obvious to the group I was speaking to (experts at the intersection of AI and Geo), but to pretty much no one else. We were about 30 people and I’m not sure many more in the UK (in relative terms) might appreciate the power of these ideas. We need to change this and I think this group is ideally placed to do so. Then I moved on to the two key challenges I see in widespread adoption of embeddings among social folks. One is definitely technical: it is still tricky and cumbersome to work with embeddings. This I’m less worried about because I can see how there are ways in which we could lower the access barrier and, more importantly, we already have vehicles (e.g., Imago and the rest of SDR-UK data services are a great example). So, things seem in motion on this front. The other one is more philosophical: embeddings are tremendously useful, but they’re not the most transparent way to work with imagery, precisely because the compression that makes them very useable also makes them obscure. For good reasons, social scientists and adjacent folks tend to be very sceptical of obscure measurements. But this doesn’t mean there’s no value at all in engaging with the technology. We need a lot of evangelism and a bit of research to bring better understanding of how to use embeddings productively in these domains.

And that was that! In classic professorial style, I totally over-run my five minutes, something I’m not proud of. Weiming very politely brought everything back in line and we moved on. Again, thank you so much for putting together such group and for thinking of me as part of it! If any of the above resonates with your, please do get in touch with us at Imago! We’d love to hear from your and start a conversation.

Here I focus on images because this is where my interests lay, but the idea of embeddings extends to pretty much any data type a neural net can deal with. Which is to say, to pretty much any data type you can think of. ↩︎

The case for satellite imagery in social science and policymaking: A sketch from my talks in China

Thu, 29 May 2025 23:26:29 +0100

Last week I came back from ten days in China on a research project. It was a whirlwind tour with five cities, and lots of opportunities to interact with both the, academic and otherwise, Chinese world. It was fascinating. At many levels, it felt like a peek into the future.

As part of the trip, I had the opportunity to give two talks, one at the Chinese Academy of Sciences (CAS) in Beijing and another at Southwest University in Chongqing. Both were titled “Making satellite imagery more useful, usable, and used in Social Science and policymaking”, and both had a very similar structure. The CAS talk was a bit longer, and it allowed me to illustrate practically many of the ideas I talked about in the first part using our ongoing EuroFab. Here I want to focus on the more conceptual bit, which aligns very much with the motivation behind Imago.

The talk tries to build the case for paying more attention and interest to satellite imagery as a source of data and information in the context of social science and policymaking. I structured this argument in four acts¹: the value proposition, the promise, the reason, and the gap.

Start with the value proposition. We’re at the confluence of three trends that make this a unique moment to look at satellite imagery with “untraditional” eyes. One, there’s never been more and better satellite(s and) imagery available. Launching stuff into space is becoming dramatically cheaper than it used to be. As Econ 101 would predict, you lower the price of something, more of that is bought, and we have more and more satellites being placed on orbit (more data). These are also benefitting from the revolution in consumer electronics of the last few decades, so the sensors on board are able to capture much more than before (better data). Second, we have better ways of extracting information from such imagery. Images contain a lot of useful information but it’s all encoded in pixels that don’t mean anything in themselves. To make something useful, you need a “magic decoder” that turns pixels into buildings, roads, land uses, etc. Well, the last fifteen years of progress in computer vision have delivered just that. From the first deep learning models of the early 2010s, into the more advanced neural nets of the end of the decade to contemporary foundation models for vision, how we teach computers to see things has undergone nothing short of a revolution. And third, which seems less relevant but it’s the piece that makes it all possible, the technology stack required to access such algorithms and operate on such data has been dramatically democratised. Compute is cheap(ish) and software is free(ish). It’s a brave new world.

Now the promise. What the confluence of the three trends above means for social researchers and policymakers is we can now do at least two things we couldn’t before. One, to measure things we haven’t been able to measure (at scale); and two, to measure better, sooner and/or more often things we’re now measuring slowly, late and sparsely. Both are, in some ways, the two sides of the same coin. They capture how the same “opening up” of new sources of information affect different status quo. In some cases (e.g., temperature, air quality, certain features of the built environment such as solar panel installations), we have not been able to measure certain phenomena that are relevant for social processes at the scale and detail required to consider them seriously in our work. In others, we do have a current solution, but it is lacking in many ways and we use it because it’s better than nothing, not because it’s good. For many of these cases, I think satellite imagery is poised to redefine what we consider acceptable for data feeding into research and policy.

Why, beyond “just because we can”, should we try to realise such promise? The short version is we’ve never needed new, timely and accurate data more than now. The world is becoming more unstable and unpredictable (e.g., climate change, spatial inequalities). We can’t “wait for the next Census” (which might not even happen in some places!) to know how society and the built environment are changing. Decisions need to be made before then and, for those to be meaningful, there’s a whole lot of understanding we need to unlock first. And, by the way, this is not a clash between traditional sources of data like the Census and newer ones like satellites. We need both of them, separately, but particularly in tandem: satellites can “stretch” the value of Censuses by, for example, providing additional context; and most of the value held by imagery relies almost entirely on our ability to relate them to existing measurements of phenomena we want, of labels. Our best shot at that “labelling” in the social realm are traditional sources.

And, finally, why aren’t we already realising this promise? This is what I called The Gap. Again, three main reasons in my view. One, imagery is big. If nothing else, there is a lot of it and it takes a lot of space on storage disks. To a point that changes how you have to think about working with it (this is the “it doesn’t fit in your laptop” problem). Two, imagery is hard. The problem is not only that managing a large volume is challenging in itself, it’s that it is also non-trivial to make something useful out of images. Something something, magic decoders, something, something. And, yes, these have become much more accessible, but it’s still the case the science underpinning all of this is tricky. And, three, imagery is different. We don’t train social scientists to work with images. Local government officers and government analysts don’t know where to start to make sense of it all. This is in contrast to other disciplines where imagery is front and centre, their bread and butter. Social sciences have, for a long time, spent all of their training allowance on other skills and techniques that are fundamentally different.

So that’s it, my 20 minutes² to convince you are up. If you’re interested in these ideas and in acting on them, stay tuned for everything we’re doing at Imago. It’s still early days so much of what we’re focused on is on getting started. Like any big ship, getting momentum takes a lot of energy and time. But once you’re going…

An unwelcome departure from my beloved “rule of three” but, in this case, I do think it was warranted ↩︎
That’s what it took me in the talk, Firefox says this should take you 6 minutes to read. ↩︎

"Geospatial AI for Land Use" (my talk at OECD's WPTI'25)

Sun, 27 Apr 2025 10:04:01 +0100

Last week, I contributed to the OECD’s 48th session of the Working Party on Territorial Indicators. Working Parties are a mostly internal affair where the different members of the OECD meet around specific topics to discuss progress, share experiences and coordinate. This is my second one, and they are a very interesting experience for an academic in that much of what is discussed is various forms of applied research, but the format and delivery is rather different than in a normal academic meeting.

I contributed a ten minute talk on geospatial AI for land use, and I thought I’d summarise here what I presented. I labelled it as an “opinionated version of our recent report on the topic”, and I stand by that definition. I tried to give an overview of the opportunities I see for (geospatial) AI to support the understanding, modelling and management of land use. Much of what I said is in the report (possibly in a slightly more eloquent and formal way), but I also mixed in some of the lessons we’ve learnt from our ongoing EuroFab project with OECD, Charles University and ESA.

To me, these are the three areas where (geospatial) AI intersects with land use:

More data
Better modelling
More intuitive interfaces

Start with data. Much of what modern AI/ML does is making data that traditionally were beyond the realm of computation, computable. Think of text, audio, or images. These are all sources of (unstructured) information that land use experts have only been able to access in a qualitative and manual way. Modern techniques such as foundation models for both text (LLMs) and image make it possible to treat these sources as quantitative data. And this is a big deal. Think of the amount of imagery that is coming from satellites, or how much data about, say, the planning system is locked in PDFs and word documents. Making these as accessible as a Census table will bring new sources on which to base evidence in a more timely fashion.

Then there’s modelling. Data in themselves are not information, insight, or knowledge. The bridge between them is often modelling. Modelling in land use is of course not new. But new advances in AI/ML are giving it a notable boost. We can both make better models with traditional data (e.g., a lot of the tree-based algorithms like random forests and XGBoost have revolutionised making predictions with structured data) and expand such models to accommodate the sources of data I mention above in a native way. In land use, this is also a pretty big advancement. Think of the full-fledged industry of land use regression models, where different outcomes (e.g., air quality, land use change) is explained as a function of land use characteristics. How that function is modelled and how land use is characterised/measured is poised to radically change in the coming years.

And then there’re interfaces. How much effect and impact the results from the modelling exercises in the previous paragraph can have is mediated by how they are presented and made available to the public. It’s 2025, so I don’t have to spend many words on how modern (Gen)AI is revolutionising human-computer interfaces. An example of how this could pan out is the chatbot interface we built on top of our DemoLand tool for exploring urban land use scenarios and that features in Section 2 of the report. This is of course not an area specific to land use, but I think land use is a good candidate to benefit particularly from this trend. The outputs from land use modelling are usually sophisticated and non-trivial to understand. At the same time, if we want them to have an effect, they need to be understood by non-technical folks such as policy makers, practitioners and the general public.

And that, pretty much, covered my ten minutes! It was fun to think of how to present results from our research to folks who are not academics but deeply care about the topic. This year, I had to participate online, but I hope I have future opportunities to come in person and meet participants over coffee breaks (or even during the timetabled cocktail!).

FutureBuild'25

Thu, 06 Mar 2025 15:53:36 +0100

Earlier this week, I participated in a [panel on the Land Use Framework](Edge Debate 179) at FutureBuild'25. This was curated by Edge, and co-panelists included Baroness Young of Old Scone (Chair), Maya Singer Hobbs, Carolyn McKenzie, and Stephen Hill. Each of us spoke for 5-7minutes and then the chair opened for questions from the floor.

In my talk, titled “Land use data to inform decision making - How it could work”, I tried to make three clear points:

Good decisions are linked to the “state” of land, and this is too complex to understand without direct evidence. (Good) data provide the context for (good) decisions.
We need data that changes “at the speed of decisions”. This context needs to reflect the current state of affairs, which changes much more rapidly than we have been able to gather data from in the past.
There’s a lot of very cool stuff happening in the world of data to support this view of the world. For example, and I may be biased here, the single most exciting development of the last ten years is the ability to make increasingly more abundant imagery from satellites computable so it can feed into these decisions.¹

Given the prompt to talk about “how it could work”, I gave a shout to our DemoLand project, as an illustration of how to combine data and AI to support land use decisions.

I closed the intervention with an afterword, two points I really wanted to make but couldn’t elegantly squeeze in the previous part:

Data (as well as the models and, eventually, digital twins they feed into) are best thought of as human augmentation rather than replacement. As much as some might like it that way, I don’t think any of this is about the robots taking over the job of the planners, it’s about super-charging planners!
Urban land is land! This is by now one of my hobby horses. Much of the land discussion (be it on modelling, understanding or affecting it through policy) is focused on “most of the land”. This makes sense to a certain extent. After all, cities are about 5% of the land. However, if we’re serious about curving emissions and building a better home for most of the population, this 5% clearly punches above its weight (e.g., about 75% emissions come from cities whom, in a country like the UK, house about 80% of the population).

This is in part why we launched Imago, to bring all this good stuff to the social sciences, public health, and policymaking arenas. ↩︎

Books read in 2024

Wed, 11 Dec 2024 01:00:00 +0100

It’s that time of the year. ¹ Spotify just dropped Wrapped. The lists of the best everything in 2024 start rolling on. And I can’t help having a go at mine. ² One of the few practices I’ve kept over the years, even in the face of pretty substantial life style changes (e.g., baby) is reading. So, here’s my list of books from 2024. Let’s get all due caveats out of the way first of all. It’s technically not a “best of” but rather “all of”. I did not love all the books in there, but I did read them. Unlike many smart people profess, I have a really hard time giving up on a book I have started. This year though, I’m glad I stuck with all of them. Also, technically, I did not read all of them either. A good chunk of them I listened to while doing the dishes (more on this below). And, at the risk of stating the obvious, it’s December 11th., the year has not ended. I’m currently on three books (one, two, three) that I might or not finnish before the year is out. All are very good. Unless something radical changes, none would replace my picks this year. So, on we go.

If by best we mean a book that makes you stay up way too late reading, and you think about long after you’ve finished, the clear winner this year is The Deluge. It’s hard to state in words how much I liked what is a 900p. book mostly about climate change. Don’t be fooled though, this is not the type of sci/cli-fi where characters and story are simple vehicles for the Big Ideas. There are lots of those (Big Ideas), but there are also real characters. People you come to love, despise, and find uncomfortably close to you. Since I finished it, I’ve been looking for someone, anyone, who’s read it too and is up for a coffee, because there are so many bits I’d like to discuss with another human about this book. Unfortunately though, the search goes on…

The other book that surprised me in many and good ways was Doppelganger. I’m the proud owner of a 1,999 paper copy of NoLogo, which I devoured as a highschool student. After that, I figured there would be little in Naomi Klein’s writing I wouldn’t agree with or would change my views, so did not read more. How wrong I was. Admittedly, this is the most and least Klein book I can think of. It is something only her could have written and only now, after decades of craft. And is a really hard book to pull off. It’s about so many things (her confusion with Naomi Wolf, conspiracy theories, right-wing politics in the XXIst Century, doppelgangers in culture and art over history, Jewish identity…) that it is hard to classify or even summarise. But it’s well worth your time. Go read (listen to) it.

Honorary mentions. This has been a good year for reads. It’s also been a terrible year in many personal aspects of my life. In the worst of it in May, I picked up Where there’s a will, and I’m so glad I did. Chappell’s account of what it really means to win a race so brutal like the Transcontinental is everything I needed to read about human endurance, empathy, and humbleness. Impossible people was also a phenomenal (graphic) read. Wertz’s account of her own alcoholism is so raw, hopeful, and just pure funny. And last but not least, Devil in a blue dress. My friend Al turned me on to this, and I’m so glad he did. This is L.A. Confidential from the African-American perspective. It’s a great detective story, it’s so well written, and it’s an approach to race that, although written in the early 90s, reads just as fresh in 2024.

A couple of thoughts to close. First, almost all non-fiction I consumed this year was listened. I’d love to have the brightness and mental capacity to spend 40 minutes every night getting lost in thorough arguments. The truth is I don’t. Instead, I do have about that time every night when I wash the dishes and clean the kitchen. One day, in an ideal world, I’ll do all my reading, well, by reading. For what is worth, I do agree with the snob argument that reading is a better way to engage with writing than listening. It gives you more space to think and let the content change you. It’s just that such option is not in the cards right now. Second, I have tried to read as much from women as from men now for a few years, with varying degrees of success. This year I think I was fairly OK with it. This year, it’s almost been a perfect split between fiction (men) and non-fiction (women). I need more female fiction recommendations! And third, something I started this year is to systematically take notes and write down impressions as I was reading. I started towards the autumn, so I’ll reserve conclusions on that for the future (including whether the practice sticks at all!).

So, that’s it, one more year. If you know me a bit, you know I like talking books, and I’m always looking for more recommendations. So feel free to hit me up with books that have made your 2024 a better tour around the sun!

Books of 2024

Non-fiction

Where There’s a Will: Hope, Grief and Endurance in a Cycle Race Across a Continent, Emily Chappell.
Impossible People. Julia Wertz.
Doppelganger, Naomi Klein. [audio]
Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence, Kate Crawford. [audio]
When the heavens went on sale, Ashlee Vance. [audio]
The making of a manager, Julie Zhuo. [audio]
The bilingual brain, Albert Costa. [audio]
Hunt, Gather, Parent: What Ancient Cultures Can Teach Us About the Lost Art of Raising Happy, Helpful Little Humans, Michaeleen Doucleff. [audio]
The Book You Wish Your Parents Had Read (and Your Children Will Be Glad That You Did), Philippa Perry. [audio]

Fiction

Spook street (Slough House, Book 4), Mick Herron.
The tusks of extintion, Ray Nayler.
Devil in a Blue Dress, Walter Mosely.
Hum, Helen Phillips.
The Deluge, Steven Markley.
The Peripheral (Jackpot #1), William Gibson.
Time Shelter, Georgi Gospodinov.

If you want to see what last time I did this looks like, head over here. ↩︎
You can see a slightly fancier version here. I’m reluctant to pay much attention to it because a) it counts my two dropped books as read and b) because… Goodreads. It’s 2024, your service is no better than in 2014. Most frustratingly, it could be sooo much better. Oh well, digression… ↩︎

Report on AI and Land Use

Thu, 14 Nov 2024 01:00:00 +0100

Geospatial AI for Land Use, by The Alan Turing Institute. Independent Report from The Alan Turing Institute - URL

How geospatial AI can help inform our land use choices. Blog post from the Geospatial Commission on the report - URL

TODAY, the Geospatial Commission released the report we¹ prepared for a project we have been working over the last year in collaboration with them. The work builds on our earlier contribution to the National Land Data Programme² last year, and the document puts in writing much of what we presented to the minister for AI and intellectual property last March. There is also an accompanying blog post Mehul and his team at the Commission put out.

Most of the report is a summary of what we learned in two specific exercises and a series of engagement events. I’ll give you the two-sentence version of each here but, if you are interested, you really should grab the report as it covers them much more comprehensively. In one, we took our initial DemoLand app, a tool that helps users explore different land use scenarios and their effect on areas such as air quality or house prices, and embedded a chat interface powered by a large language model. The original tool provided access to a lot of data and modelling that would be hard to access for non-technical audiences otherwise, but it still required the user to “know their way around”. With the new chat-based interface, exploring the results is a much more conversational experience that can reach larger even audiences. In the second exercise, we digged into the guts of the models that power land use applications such as those in DemoLand. Typically, many of these require data that is collected and released more slowly than ideally required for decision making (e.g., Census or building cadasters). We explored complementing, or even replacing, these by widely available satellite imagery. Satellite data is definitely not a new thing, we’ve had metal boxes orbitting the Earth since at least the 1950s, but there are a few recent things that make them more appealing. The revolution in computer vision (and what is now also termed AI) we have seen in the last 15 years has changed what we are able to do with imagery, even that of limited spatial resolution. In this exercise, we explore foundation models for vision (a bit like the GPT in ChatGPT but for satellite data) to see how far we could push them. The result is a series of what I consider extremely exciting results but others might see as largely dry performance scores tables. You can get a summary of those in the report. In addition to these two exercises, we also ran consultation events (at AI UK'24 and online) with experts to get a broader view on the potential, challenges and immediate opportunities of geospatial AI for land use. I could summarise those, but I won’t do a better job than the report, so head over to its “Section 3” for that.

What will probably (I hope!) catch most attention is the “Recommendations” section. Here is where we brought together everything we learnt in these exercises to propose concrete steps forward. In particular, we mention five:

1/ Identify additional areas of opportunity for satellite data to build the value case for geospatial AI.

2/ Develop a Geospatial AI Toolkit for LLMs.

3/ Expand the conversation on national foundation models to land use and geospatial.

4/ Improve access to key computational and data resources.

5/ Promote knowledge sharing and cross-discipline collaboration.

Some are self-explanatory and I suspect few will disagree with them (who is against more knowledge sharing?). Others bring to the front discussions that we think deserve more attention than they’re currently receiving. LLMs, for example, are not very good at geography (there is a reason why the second L is not a G!). Before we jump in and take them off-the-self, we think there is work to do to develop the “Geography curriculum” we’d like these models to know when they help folks on spatial domains. And others seem more obvious than they actually are. Suggesting in 2024 that satellite data be used for land use change may cause unreparable eye-rolls among land use experts who’ve been doing this in an academic context for several decades. Yet there is still very little of it that has made it into “production” at scale, particularly in non environmental and physical domains such as cities and society³. So think twice before sending your eyes upwards.

Above all, and ironically given that it summarises past work, we hope this becomes a conversation starter. There is so much exciting stuff that is happening in the world of AI that could have tremendous implications for land use modelling and management. What exactly and, specially, how are debates we have just begun and are far from done. But we think they are worth having and actioning on. To work.

The team on this project spanned folks at the Commission and Urban Analytics@Turing. At Turing, Barbara Metzler and Stuart Lynn did most of the heavy lifting, and I was the lucky one to “see early demos and present it around”. ↩︎
Geospatial Commission (2023). Finding common ground: Integrating data, science and innovation for better use of land - URL ↩︎
This is one of the core ideas that made Imago a winning proposition, so watch this space. ↩︎