Too Many Pixels...

There just too many pixels falling from Earth Observation (EO) satellites…and no one to buy them. First, a shout out: Bravo to Joe Morrison (@mouthofmorrison) from Azavea for the most insightful article I’ve read on the absurdity of trying to buy imagery from one of the commercial EO satellite companies. “The experience of buying satellite imagery is still so shitty it feels intentionally awful… It’s closer to a hostage negotiation than a sales transaction,” said Morrison. Brilliant.

There are two issues today regarding the commercial EO business:

1. The ease of purchasing EO data (if you are not a government client) as Morrison states well.
2. The information from EO data that is commercially viable, useful and provides the intrinsic value to your business application.

CONTEXT AND HISTORY OF THE EO BUSINESS

Let me provide some context. I started my career as a remote sensing geologist (today you’d call me a ‘data scientist’) with the U.S. Geological Survey’s Earth Resource Observation and Science (EROS) Center in Sioux Falls, South Dakota. EROS was and still is the place where Landsat data is archived. My job was processing Landsat-1 imagery and extracting geological data and information from millions of pixels. Later, I worked for an oil exploration company doing the same. At that time in the early 80’s when I started, if you were not a government agency (i.e. EROS, USGS, etc.), an oil company, or a university with a huge grant, you would not be able to afford satellite data. Period. Fast forward to today, if you are not the military or an intelligence agency (e.g. NGA, NRO), Morrison rightly articulates that Maxar, Planet and Airbus are not incentivized or perhaps incapable of selling any interested party their satellite imagery because of the large government contracts that constitute the vast majority of their revenue. Not that they don’t want to, but as Morison suggests, they just don’t want to put the resources into doing it with a great deal of customer satisfaction in mind.

CONSIDER WHAT YOU REALLY NEED

The solution to #1 above is a Data-as-a-Service (DaaS), a cloud solution, capable of tasking, pricing, and selling only the imagery you need from the area you need it. Both Planet and Maxar offer a DaaS interface but I find both difficult to use and navigate. More importantly, they assume that you know something about EO imagery and the variability of the data that is available for purchase. This is no small task. Let me give you a personal example: During my Master’s thesis defense, I was asked to recite and explain an image classification algorithm that I used for image processing, i.e. “the math.” That was just not my thing. I was a geologist; I could care less about “the math.” My thesis was on mapping. I understood how to use image processing software, the input data needed for creating an image-map and a fair amount about the reflected electromagnetic radiance values obtained by scanning sensors onboard Landsat. My point is that you need to know more than you think in order to make the right purchase. I’ll explain more in the example below.

The solution to #2, above, is also not so simple but ties into #1. If you actually know what kind of information you want to derive from EO imagery and its value, then maybe perseverance to navigate #1 will be worth it.

There are now over 100 commercial EO satellites in orbit besides those launched by the U.S. (e.g. India, Europe, Japan, South Korea, UAE). Knowing the required spatial resolution (i.e. pixel size and the area it represents on the Earth's surface), and spectral bands (e.g. visible, near IR and the specific band width) of the electromagnetic spectrum needed, plus the image processing software that is a good fit for your use case, will only be identified once you determine the information you seek.

So, first, why on Earth do you need satellite data? What are you looking to find? Commercial applications, that is, the use cases needed for applications in precision agriculture, insurance risk analysis, or retail real estate development, vary widely, for example. Do you really need satellite data at all? Will drone imagery suffice? Fixed-wing platform photography? When do you need the data? Is your need for information ephemeral? Does the combination of spectral and spatial resolution match your use case?

YOU NEED THE ANSWER

Who cares? What you want is the ANSWER! Do you expect the pixels to arrange themselves neatly into something you’ll comprehend? What you need, and expect, are the areas of land needing an increase in irrigation. You need to know which houses were impacted by the hurricane and the policies written to cover property damage. You need to know which areas have greater than 500K square feet of developable land in areas zoned for commercial real estate. You want the building footprints exceeding 10 stories in height.

Those are answers. When some entrepreneurial company figures this out, it might not be imagery that will be sold at all. It will, however, be the information derived from the imagery that provides value. This information may be from satellite imagery; or it could be from a drone tasked to fly a specific region. It might even be an archived image because you don’t need data that were necessarily captured within the last year. You really don’t care. And, you will not want the average knowledge worker in your company to figure this out.

Herein lies the promise of machine learning; the ability to have image processing algorithms learn and identify classes of objects, e.g. the number of cars in a parking lot, geologic rock and mineral formations, the size and extent of the flood zone, etc. Users would input the object they seek; the algorithm seeks available imagery and returns the object. [My preferred example of a spatial query that would utilize ML and NLP is to ask the following: "Find me all of the Class A office space available for rent in a two square mile area between 43rd Street and 7th Avenue in Manhattan where the traffic count averages 20,000 vehicles between the hours of 10 AM and 2 PM."]

Briefly, as an example of what is described herein, let’s look at the above image-map created after the explosion in Beirut, Lebanon caused by fire in a warehouse holding ammonium nitrate on August 4, 2020. The article describing the image above notes the source: "NASA Earth Observatory image by Joshua Stevens, using modified Copernicus Sentinel data (2020) processed by ESA (the European Space Agency) and analyzed by Earth Observatory of Singapore (EOS) in collaboration with NASA-JPL and Caltech, Landsat data from the U.S. Geological Survey, and data from OpenStreetMap." And further, “The team at ARIA (Advanced Rapid Imaging and Analysis team at NASA Jet Propulsion Laboratory) and EOS examined synthetic aperture radar (SAR) data collected before and after the explosion, mapping changes in the land surface and built structures." Finally, the map legend: "Dark red pixels represent the most severe damage, while orange and yellow areas are moderately or partially damaged. Each colored pixel represents an area of 30 meters by 30 meters."

Did you get all that? My point is that if it takes this much to create the map, do you bother? Further, while I very much applaud the effort, the map above is not revealing the entire answer either. It doesn’t tell me which buildings incurred the most monetary damage or comprised the most impact to the population in terms of lives and business locations. This would be critical for government agencies to assess recovery and rebuilding efforts, just to name a few use cases.

To state the incredibly obvious, acquiring and understanding EO imagery data needs to be easier. As Morrison points out about the satellite imagery providers, “You’re sitting on one of the greatest underutilized digital assets in history.” We are living in the information age, but right now there are just “too many pixels.”