Nobody wants your fancy algorithm
The satellite imagery industry still has no idea what customers actually want
Disclaimer: all dumb opinions expressed in this newsletter are mine. Any good ones are properly cited with who I stole them from. I am once again asking you to just…be cool. Just be cool. Thank you in advance for not getting me fired because of something I write here.
This is how I would summarize the commercial satellite imagery industry today: a bunch of raccoons (defense contractors) dressed up in trench coats (Silicon Valley startups) pretending to be humans (SaaS products).
When I see a company trying to turn commercial satellite imagery into a one-size-fits-all derived data product and selling it as a subscription product, it feels like greeting an old friend. He’s still down on his luck. He thinks this time is gonna be different. Oh no…he’s gonna ask me for money again, isn’t he?
When it’s a competitor falling into this trap, part of me kind of enjoys it…
…but most of me feels frustrated. After all, I am not rooting for people to fail. We’re building an industry together, not playing a zero-sum game.
As always, I could be totally wrong. Only time will tell.
Error 404: Customer Not Found
Allow me to put it more succinctly: selling derived data as a subscription product does. not. work. I don’t care what it is. The juice is never worth the squeeze.¹
Count cars. Count airplanes. Count ships. Segment land cover. Track oil inventories. Estimate biofuels. Measure water levels. Etc. Etc. Etc.
Now! You can give that data away for free and people may use it—in fact, I think derived data streams (“information feeds”) could be a force multiplier for accelerating realization of the value trapped in massive imagery archives. There are many notable examples of derived datasets being given away for free to much fanfare.²
But sell derived data feeds? Straight up? As a subscription?
Even if it did work (it doesn’t), I’ve written before about how satellite imagery companies entering the analytics business is not my favorite thing in the world in general. There are many reasons that I hold this opinion, but principally:
It’s hard to be a great analytics provider, harder yet to be a great satellite imagery provider, and nearly impossible to be great at both simultaneously
You wind up competing with your customers, and that’s, like, not the vibe 💅
Satellite imagery is inherently applicable to many problems, so you’re fighting gravity to try to focus on just one while getting pulled in a thousand other directions
But it gets worse, because satellite imagery companies are not the only offenders here. In fact, I see lots of new analytics startups pursuing the dangerously seductive strategy of building a new type of dataset and then closing their eyes and hoping customers will magically show up to start buying it.
“Tell me,” I whisper to them, “The customers who want your derived data feed…are they in the room with us right now?”
“No one wants pixels, they want insights” is wrong
There’s an almost ubiquitous cliché in the satellite imagery business that goes something like, “Nobody wants pixels, they want insights.”
It sounds really wise. After all, 99% of the people who can benefit from satellite imagery do not need to ever see it directly to receive that benefit (and wouldn’t even know how to open the file if you sent it to them).
It turns out some people do just want pixels and they spend hundreds of millions of dollars a year on satellite data, so that part of the aphorism is wrong.
Even worse, the second part is wrong, too. It turns out nobody wants insights, either!
What people want is for you to solve their problems for them. And people’s problems aren’t magically solved when you confront them with an “insight.” There’s still action that needs to happen as a result of that insight for value to be created.
Actions are facilitated by applications. I use applications to get real work done every day:
Using Slack to send YAKOTHGIF (yet another King of the Hill GIF) to a coworker
Using Spotify to spend half an hour finding the right playlist to blast while I fill out an expense report for three minutes
Using Excel to shade in the cells so it makes a big smiley face
Without an application, a data feed often winds up in a bad spot: you can produce the right information at the right time with no way to monetize it.
There’s no such thing as a generic algorithm
Ok, so the main reason not to launch data feed products is that there is no demand for them, and therefore by definition they are a waste of time and resources (unless you’re trying to release the data for free in order to help the world or grow the industry or whatever).
But even if there were demand, you run into a second insidious problem. It turns out everything related to satellite imagery analytics is harder than it sounds. Everything. If you want examples, I’ve linked to some relevant blog posts in the endnotes.³
In my opinion, every supervised machine learning model is hopelessly biased by the intent of its creator(s). Namely, it inherits the bias of its training dataset (both geographic and semantic).
Unfortunately, the world does not neatly fit into a clean, universal taxonomy—at least not one that is specific enough to be practical in every day matters. So, you have to make editorial decisions. And that creates bias.
The most exaggerated form of delusion about generic algorithms is when companies release “change detection” products as stand-alone, all-purpose services. What constitutes meaningful change? You can’t answer that in advance of understanding a particular customer’s needs, and therefore you can’t train a supervised model to detect it a priori.
The most practical approaches I’ve seen to subscription data feeds are always custom. You can “fine tune” a generic “pre-trained” model with a relatively small number of examples (1-10K) and get incredible results this way. It’s a compromise, and it certainly does not produce 90% margin SaaS-like revenue, but at least it works and some people will pay for it.
An alternative (and far better) business model
Here’s the thing—the people launching data products are on the path to success. They’ve just pulled off at a rest stop thinking it was their final destination.
Underpinning ALL of my favorite companies in this industry is a custom data feed. Arturo has an algorithm that can predict relevant property characteristics for insurers using satellite or aerial imagery. Cloud to Street can predict flood risk and map flood depth using synthetic aperture radar from space. NCX can offer an efficient marketplace for carbon credits by verifying what suppliers tell them using their pre-computed dataset of carbon stocks. Upstream Tech uses water levels inferred from satellite imagery to predict hydropower generation.
What’s different about these companies? They didn’t stop at the underlying data feed. They build applications that their customers use to take actions and solve problems. If you use an application every day, you don’t mind paying a subscription for it.
Even satellite imagery companies can take this approach. It requires laser-focus on a single industry, which is quite difficult for satellite companies to do. On the whole, we’re a greedy bunch—we tend to want to have our cake and eat it, too.
The problem with data feeds is that they’re easier to work with than raw data, but not by enough. Anyone savvy enough to build an economic indicator application on top of your ship detection algorithm is very likely to also be savvy enough to just… build their own ship detection algorithm (and optimize it for their needs over the idealized generic model).
Plus, application companies like to build their own algorithms/data feeds because they view it as part of their proprietary advantage—it’s incremental investment that yields asymmetric defensibility.
A big industry with a handful of customers
I believe an underlying assumption that motivates people to keep launching these ill-fated data feed products is that they (rightly) believe there is massive, untapped potential in commercial satellite data and the industry is set to grow quite quickly as we collectively find ways to unlock that value.
I can understand the leap in logic from acknowledging not many people work with satellite data today to reasoning that launching new products that are more refined should lower the barrier to entry for new customers. More customers = growth.
However, if you believe the revised credo, “customers don’t want insights, they want applications,” then you might come to the same conclusion as me: the satellite imagery industry will always be small. Only a handful of companies will ever solve their own problems with custom applications built on data feeds—the rest will happily pay someone else to just solve the problem for them.
Crucially, I don’t mean commercial satellite imagery always be a small industry in terms of the dollars that flow through it—just small in terms of the number of firms that constitute it.⁴
¹ Ok, ok, there are exceptions to this rule. VRICON comes to mind - they spun out from DigitalGlobe, made a derived dataset of global elevation data, and then were re-acquired by the same entity that spun them out at the first opportunity. I think they have one huge customer in the U.S. Army and that on its own is enough! Ursa Space’s oil inventory product may be another major exception; they outlasted Orbital Insight which used to have a competing product but eventually shuttered it. However, these are notable exceptions to a simple rule that holds up surprisingly well to the many, many varied attempts at this business model over the last 20 years.
² Global Forest Watch is my favorite (forest cover loss). California Forest Observatory is another phenomenal example powered by my friends at Salo Sciences in partnership with Planet. I’m also partial to the USFS Wildfire Risk app that my former colleagues at Azavea made along with Headwaters Economics. I could go on, and on…Esri/Impact Observatory land cover data, Meta (Facebook) Global Population Estimate, and on, and on…
³ I’ve written up a bunch of examples of how a seemingly simple task like “count cars” becomes excruciatingly complex very quickly in practice. Is a truck a car? Ok, then is a food truck a car? Ok, maybe we should have said, “vehicles.” Cool, then is a boat a vehicle? Is a bicycle? No, ok. Well, then is a motorcycle? And on, and on, and on. And if that weren’t hard enough, the misdirection carries all the way down into the very “quantitative” metrics we use to describe “accuracy” which are just as opinionated as the models themselves. It’s a mess.
⁴ For a sense of what these firms tend to look like, here’s a thread on the topic: