The impact of high-res Earth observation data on AI and data downlink
In this episode, Claude Rousseau, Research Director at Analysys Mason and expert in space and satellite, speaks with Peter Wilczynski, Chief Product Officer, Maxar Intelligence. They discuss the demand for Earth observation (EO) very-high-resolution (VHR) imagery and its impact on data downlink and artificial intelligence (AI)/machine learning (ML).
Peter shares his knowledge and perspectives on:
- the value proposition for VHR imagery and the specific uses cases that this can enable
- the impact of VHR imagery on data downlink with regards to latency and data rates
- the downstream analytics solutions that have developed as a result of VHR, and Maxar’s role in the downstream market
- the impact and role of AI/ML in EO products and services, and the associated benefits for internal efficiency and customer-facing applications
- how the EO market will move away from being a niche, government/military-focused market to being a more-enterprise-focused space.
Tune in to gain valuable insights and learn more about very-high-resolution imagery and how it is reshaping the Earth observation data downlink and AI markets.
Transcript
Claude Rousseau
Welcome to the Analysys Mason podcast. My name is Claude Rousseau, and I am the Research
Director for the Space Practice at Analysys Mason. Today, I'll be discussing the demand for EO very high-resolution imagery and its impact on data downlink and AI/ML. I am joined by Peter Wilczynski, Chief Product Officer at Maxar Intelligence.
Hello, Peter.
Peter Wilczynski
Hi, Claude. Thanks so much for having me.
Claude Rousseau
My pleasure. Can you just tell us a bit about you and how the topic of very high resolution, as well as AI/ML, fits into what you're doing with Maxar Intelligence?
Peter Wilczynski
Yes, thanks so much. So, I've been working in mapping geospatial intelligence for most of my career. I spent 12 years at Palantir and recently came over to Maxar. I have always been really passionate about the role that imagery can play in taking a lot of different data sets and grounding them in a real understanding of the world. When I think about data analysis, so much of what we're doing is fundamentally geospatial. I think high-resolution imagery and imagery generally has a very big role to play in grounding our analysis.
Exploring very high-resolution imagery
Claude Rousseau
That's great. Thanks for that introduction. Well, let's jump right into the topic. Could you share with us some insights about the value proposition for very high resolution - which we talk about, as below 50 cm ground, and high resolution, which is normally between 50 cm and one-metre imagery?
What are the specific use cases for this that you can address today that were not possible before?
Peter Wilczynski
Yes, so, stepping back. Maxar has been a provider of very high-resolution imagery for a long time. We brought some of the first commercial imagery to market with the launch of early satellites in the late 1990s and have been going ever since.
Today, we run the most advanced set of on-orbit capabilities for collecting very high-resolution imagery and then use that imagery to produce a ton of downstream, very valuable derivative products off of the imagery. So, when I think about imagery, there are two key attributes that make imagery valuable. It's the resolution and the accuracy. And they're often really related.
So when you think about the core foundational mapping mission that Maxar serves, helping commercial and government organisations create maps of the entire world, high-resolution and very high-resolution imagery is really a must-have. To distinguish between different buildings, to distinguish between different pieces of road furniture. It's always been really critical to have exquisite imagery as a base layer for a global map.
What I'm excited about today and moving forward into the future is that it's really critical for machine learning applications. That's because you have a lot more data, you have a lot more raw information, for lack of a better word, to feed into AI models and extract insights and information.
As humans, we look at the world, and we bring a lot of context to it. We know what a road looks like from the ground, we’ve driven on roads, we've walked alongside roads, we've looked at buildings, and we understand a rich amount of information about our spatial world that we bring to bear when we look at overhead imagery. But a machine learning model doesn't have that, and so you really need to put as much information as possible into that image so the model is able to build that same spatial reasoning into itself. And so we see, especially for feature extraction and fingerprinting workflows, understanding specific buildings, specific ships, differences between left turn lanes and right turn lanes, dividers versus medians, all of these things are really critical to be able to extract from an image. For a machine learning model, they need that high resolution to get that information.
Claude Rousseau
Would you say, therefore, that's enabled a few new use cases today that, say, five years ago, you wouldn't have imagined working on?
Peter Wilczynski
Yes, that's absolutely right. You know, we've been working with machine learning as part of our production pipelines for 10 to 15 years. We have customers and partners who do really advanced production techniques on top of the imagery and have done so for a long time.
When we think about machine learning at Maxar, we think about three things: data, algorithms, and compute. I think the thing that everyone's excited about today is an advance in the algorithms. But from Maxar's perspective, we're really focused on providing high-quality data for those algorithms to use. The key workflows that we're seeing improvements on are not necessarily the kind of problems we're trying to tackle. We've been doing segmentation and object detection for a very long time. It's just the performance level of a lot of these new algorithms that they're able to hit is much higher than five or ten years ago.
Demand for high-resolution imagery
Claude Rousseau
That performance, I guess, does, therefore, impact the demand for some types of imagery. Which one do you see as the most in-demand today? Or which one has the biggest growth rate between high resolution and very high resolution?
Peter Wilczynski
We're very focused on very high resolution at Maxar. Part of that is because we see it as foundational for any other workflow that you're going to do. A lot of our work is about improving imagery generally across the ecosystem, whether that's by co-registering it with our imagery so that you can get higher accuracy from other providers or by doing HD enhancements.
So, taking 30 cm imagery and synthetically generating 15 cm imagery is a key part of one of our Vivid Advanced products. The nice thing about very high-resolution imagery is that it allows you to essentially baseline everything else off of it, co-register it into foundational content, and then drive more value by essentially enhancing even high-resolution imagery by co-registering it. So, we see the general demand for optical imagery continuing to grow.
For a long time, we've been capacity-constrained. We've launched four new satellites this year and are able to deliver another 2 million square kilometres a day, which is a really big improvement for our customers. So, yes, we see the AI/ML desire driving demand, but also just increasing utilisation of spatial data driving demand.
Claude Rousseau
So, you just said you've launched more satellites, and that probably brings you to having more data to downlink from those satellites. Given that those satellites also provide you with higher resolution information, imagery and information, has that affected how you downlink data today?
Innovations in data downlink
Peter Wilczynski
Yes. The entire pipeline from space all the way down to the ground is incredibly sophisticated at Maxar. It starts with space-based compression, taking the raw pixels that we're getting off of the sensor and compressing them before we do downlink, caching them prior to getting above-ground station, and then downlinking directly.
We really focus on end-to-end latency as a core metric that we're trying to improve from an innovation perspective. We have a number of unique offerings for placing downlink stations at customer sites directly via our direct access facilities. So that's allowing customers to cut out the latency of a satellite orbiting over our own downlink facility or one of our partner downlink facilities with AWS and KSAT to get a much better latency reduction.
When I think about network topology, it's really about bandwidth and latency, and I think we see really big improvements moving forward. What are next-generation satellites? We just launched satellites, and we're already thinking about the next generation and how those are going to be able to take advantage of some of the new space-based mesh networks that are going up and get data to end users more quickly.
But downlink's really only one part of the puzzle. It's a pretty important part, but the other piece we're focused on is our production systems on the ground and how we reduce latency there. So a lot of the re-architecture that we're doing internally is moving from batch production to more event-driven streaming production, where, as you make these production systems more sophisticated, you add more derivative data products that you're trying to produce, you want to be able to do more just-in-time production of individual data products based on customer demand and customer orders, versus producing everything up front, regardless of who wants it and when they need it.
Advancements in analytics
Claude Rousseau
Well, that leads me to my next question: what you're talking about, batch production and event production, I guess somewhere down the line, you're also improving your analytics. Your analytics solutions are developing as well, as a result of very high-resolution data being available. Could you expand a bit more on that? And how has your role at Maxar changed over the years in the downstream market?
Peter Wilczynski
Yes. When we think about analytics, we think about two distinct transformations. One is for a given image, extracting vector information from raster information. That's about automating imagery exploitation, so extracting land cover information, extracting buildings, roads, cars, vessels, and trying to understand the physical state of the scene that the image is describing. Then, also understand what's happening at any given time. So that's taking one image and extracting it out, making sense of that individual image. But where we're more focused on from an analytics perspective is really how we think about deep image stacks. So, moving from single image analysis to thinking almost like a movie of many images. A time series of information. It involves extracting out geospatial features that represent semantically meaningful segmentations of the world. So, parking lots associated with a particular restaurant or mining sites associated with a particular supply chain. Then, build out time series information so that you can really understand the change over time of key indicators. Whether that's the volumetric size of iron ore heaps at a mine or the count of vehicles at a particular production factory. When we think about analytics, we think about trying to tie together these sorts of spatial features with a more object-oriented analytical process where users and end-user analysts are thinking in terms of objects and relationships between objects and moving away from just purely looking at pictures and trying to understand what's happening from just a pure picture perspective.
The role of AI/ML in operations
Claude Rousseau
It seems that, in that instance, what you're doing is also bringing more new, enabling technologies like AI/ML to get your EO products and services to the market. Has that been impactful for you so far, or are you still a work in progress?
Peter Wilczynski
Yes, so we use AI/ML technologies throughout our whole production stack today. And when we think about moving forward, it's really about continuing to embed them in end-to-end workflows. When we think about this for our customers and for ourselves, we think about there being three categories of places where we can apply AI/ML to our system. That's in the operations, so mission planning, scheduling, simulating the constellation, understanding where to take images and how to balance supply and demand for particular regions. How to staff our teams, and how to make sure that the satellites are staying operationally healthy. That's a really rich vein of applied ML.
In our production systems, which are taking raw imagery and producing finished products, we already use a lot of AI/ML today for things like atmospheric compensation, cloud cover reduction, and imagery enhancement. That's a key part of how we deliver very high-resolution products today that are extremely accurate and extremely easy to use for end analysts.
The last piece where a lot of folks are spending time right now is on image exploitation and analysis. So, being able to take those finished products and create insights out of them, create intelligence out of them, working as a co-pilot to a human analyst, trying to understand what's happening in the world. So, yes, we see really big applications for AI/ML across those three tiers of operations, production, and exploitation. I think moving forward, we're really excited about how we connect the loop from analysis back to collection. Helping to drive towards more autonomous tasking, autonomous collection management, and understanding how we drive more efficient orchestration of a larger and larger constellation. Not just of our satellites but also partner satellites who may have different capabilities and bring different phenomenologies to bear against some of the same analytical problems.
Claude Rousseau
You just mentioned the word partners. So, I would ask you then to explain how you see the adoption of AI/ML across the industry. Do you feel that it's taking it on like wildfire, or is it more like a step-by-step approach?
Peter Wilczynski
It's probably both. I think it's been moving step by step. I think the thing that's really captured people's imaginations with this moment in AI/ML is the promise of things that were just, frankly, not possible five years ago. I think that's driving a lot of excitement and a lot of ambition in terms of how far people want to take AI/ML. For the first time, people are starting to reimagine, what if I was able to do automatic scene understanding? What if I was able to interact visually, verbally with images, with individual scenes? Put together analytic reports. You know, do painstaking geospatial workflows in a much lighter-weight way? I think that's capturing a lot of people's imagination, and I think that's the part that's really spreading like wildfire. The potential of this moment that we're in.
I think in terms of the implementation into production, that's what's moving forward a little bit more step by step. These things are really challenging to develop, and I think operationalising them requires a level of orchestration that isn't just about developing a new AI model but also chaining that model together with other models piecing that together with other data systems. Integrating that, harmonising those data sets, really understanding how to normalise the representation of the world into a consistent data representation, and then feeding that to even more models.
I think we've been focused on building individual models, but I think where we're going to see a big step change for the end user and for the industry is these models of models or systems of systems, where the output of one model is being fed into the input of another model. So, going back to that example earlier, maybe you have one model that's doing object detection, and then you have a database of object detections over time that's creating a time series. Now you have a time series model that's sort of analysing causality between some sort of supply chain and a downstream factory production schedule. Those sorts of models are really playing together in a pretty symbiotic way, but each piece has to be developed independently, and they all depend on one another.
Shifting market focus to enterprises
Claude Rousseau
You do mention things such as implementation difficulties with that. But overall, do you see the commercial market, the EO market, changing in the next few years?
I'd like to have your views on how this market, which is mainly government military focus, will change into more of an enterprise focus, or at least add more enterprise focus to its commercial markets.
Peter Wilczynski
Yes. So, the government has been the key driver of the geospatial industry for the entire history of the industry, and I think that's probably not going to change dramatically. The government is going to remain a huge market, and it’s a huge part of the U.S. and international economy. So, I think it's more of an and than an or, in my opinion. But we do see a lot of growth in the commercial segment, the enterprise segment. That's really in two categories.
The enterprise mapmaking space, companies like Here, Apple, and Microsoft are building canonical base layers of vector maps that other people build on top of that fuel location services and fuel workflows like logistics, supply chain routing, and scheduling. Those are the core sort of foundational consumers of the EO market. I think those consumers are going to continue to build an ecosystem of location-enabled services on top of them over the next few years.
We see increased demand from the foundational map makers to have more currency, more accuracy, and maps that are more reflective of the real world. So I think that means more satellites and more imagery because ultimately, as the world changes more quickly, as buildings get built more rapidly, the change needs to be reflected in those foundational maps.
But I think the more interesting space of growth for the industry is in the more vertical-specific areas. So places like telecom, energy, mining, large-scale, industrial facility development and infrastructure projects. That's a place where I think the industry has to do a better job of meeting customers where they are. One of the things that's unique about the government market is the government has really talented, sophisticated geospatial analysts on staff, very comfortable with working with large-scale imagery, doing geospatial intelligence workflows and collaborating on those workflows. At a mining company, that's not something that is necessarily true, certainly not at a telco company. So when we think about meeting the broader market, I think we have to be comfortable delivering end-to-end solutions that use the raw pixel or raw imagery data that we're providing but do something much more specific. When I think about this, I think about what drove the demand for gasoline. It wasn't Rockefeller producing a bunch of gasoline; it was Ford producing cars. People want finished products that consume the imagery data and do useful things for them. But in these more niche verticals, it's incumbent on us as an ecosystem within the geospatial industry to produce those end-to-end solutions as opposed to trying to ask our partners to do that on their own.
Claude Rousseau
Well, those are great last words. Thank you very much, Peter, for your time today.
There is more content on this topic on the Analysys Mason website. To automatically receive future episodes, please subscribe to the Analysys Mason podcast.
Thank you once again for listening.
Related items
Article
From pixels to insights: AI/ML is transforming the Earth observation big data market
Company profile
e-GEOS: satellite operator, data reseller and service provider
Survey report
Earth observation worldwide: satellite operator survey