The problem with pixels*
Pixels are a pain and the source of all your EO product problems. Don't sell pixels; sell change.
Pixels are the atomic units of a fundamental data structure, the array. Arrays are everywhere in computing, from screens to spreadsheets to Instagram filters; arrays and array math surround us. Nevertheless, in Earth Observation (EO), the pixel is a problem. This problem is muddied by the central notion of “what is the product” of Earth observation; spoiler alert: it is not the pixels anymore.
What’s the situation(al awareness)?
Traditionally, the product of EO has been situational awareness or one-off scientific analysis. EO provides a tremendous strategic advantage for those organizations interested in the mapping and movement of people or goods either at scale or in remote (politically or physically) locations. Situational awareness is still a hugely valuable part of the EO market, and humans are well-tuned to consume this product. In the same way that we all can easily interpret Instagram posts and draw meaning from a wildly diverse set of images, image analysts with additional training can quickly review an image and account for numerous data inconsistencies to draw conclusions from an EO image.
Our brains can easily account for different lighting conditions, colour balances, spatial scales, and even some geographic inconsistencies. We are excellent multi-scale pattern-matchers. That, combined with a capacity for abstract thought and some cultural awareness, means we can infer much about human behaviour from images.
Pixels are a wonderful source of information for subjects important enough to employ human analysts. However, this activity is becoming increasingly expensive, and data flow since around 2010 has increased enormously. This means we have lost some of the strategic advantages of EO imagery. That said, those geospatial teams still exist in large geospatial institutions, and with increased compute, the speed of imagery analysis has increased. The point is that in traditional geospatial institutions, geospatial teams can receive a flow of pixels and know what to do with them; dedicated talent is available.
As discussed, with the rise of commercial space, Low Earth Orbit (LEO) has become more accessible. Whether this is purely a factor of price or payload availability is probably up for debate. But the net effect is we have many more commercial sensors in LEO than ever. Too many sensors for the number of eyes to keep up.
Timing is everything
But that doesn’t necessarily matter. Because the flood of sensory capability is coming from a breed of sensors which are less well designed for situational awareness, and better designed for monitoring. these sensors might not have exquisite spatial resolution or optical quality. A bigger problem is that they are at once everywhere and nowhere. They have a high temporal cadence or resolution, but they are not being pointed at a specific place at a specific time. This paragraph quickly draws us into a debate about camera specifications. I want to avoid that rabbit hole; suffice it to say that there are different sensors which are suitable for different purposes. Discussions on the commoditization of imagery often miss the fact that pixels from different sensors support different products and purposes. In this regard, all pixels are created equal, but some are more equal than others**. Monitoring is a relatively new and poorly understood capability.
The word “monitor” implies a context and a status, both of which are generally missing from a satellite camera's most basic output: pixels. However, with the defining feature of repeated imagery, clearly, the ultimate product is not the pixels themselves but the change measured between those pixels captured from the same location at different times. The central vector of geography has then become time.
The problem with time and change is that they demand context and management. Therefore, EO has become a problem of software, not hardware. This point is often missed by those motivated to build their future space companies and obsessing over the potential unit economics of massive data capture. The unit economics are great, but only if you understand your product. If you are selling imagery from a constellation which is not tasked, then your product is not situational awareness; it’s monitoring. If your product is monitoring, then you cannot just sell pixels; you must do much more.
Three or Four body problems
As suggested above, EO is often conflated into a single product offering: imagery. That conflation creates huge information gaps and scale effects, and it is dangerous for our community of practice. On each sensor platform, we are dealing with numerous resolutions of time, space, spectra, behaviour, storage, and compute. These differing attributes result in entirely different downstream products suitable for entirely different markets.
What is common about these markets, though, is a lack of the geospatial teams which are critical in the situational awareness product described above. What is more fearsome is that we must rely on more automation and machine- or deep-learning-based algorithms for the monitoring product. This ultimately will demand much more industry standardization, making me bring up the much-maligned subject of “Analysis Ready Data.” I am happy to be corrected, but I don’t imagine many other communities demand an extra step to make data analysis ready. It’s a tacit admission that we don’t have robust industry standards.
Double Middle
My final observation for today. If we agree that pixels are hard for non-experts to use, then we must rely on market intermediaries to build "industry consumable analytics.“ That makes sense to me; domain experts team up with geospatial teams to build domain-focused applications. However, given the rise of imagery marketplaces, which have arisen to support convenient imagery access from multiple vendors, we have created a double intermediary market. The implication is that pixels must travel through various hands before ultimately providing value to a particular user. While, on the one hand, this feels a little like ad tech, the practical movement of data is in and of itself ludicrous and wildly expensive. Therefore, there is probably a market opportunity to serve commercial data on behalf of others to avoid needless data downloading. This service needs to happen close to the marketplace and be persistent, like the photos app on your phone.
* With apologies to our friends with active sensors…
** Hopefully, you will all pick up my reference to Orwell’s Animal Farm. My point here is that it may appear that pixels are the same, but the fact is they are different, and they ultimately get treated differently. Some pixels are more valuable than others, and some have likely never been looked at by a human eye - see the lament for the lonely pixel.
Great insight Tom. I might argue that geospatial analysis has always been about change detection. Feature detection at its core is change detection over spatial dimensions. Temporal change adds complexity, particularly if spatial resolutions vary. So, I guess I would contend that central vector is a composite of space and time. Just focusing on time - assumes the feature is stationary - which is not true for features such as hurricanes, tornadoes, coastal inundation (yes I work at NOAA😬).
Kim at GeoCento has some interesting thoughts on the Double Middle, as you describe it… Great piece!