Orbital Insight (www.orbitalinsight.com) is tackling some tough market intelligence problems by analyzing satellite imagery and applying machine learning to big data to glean new insights. The company recently closed a Series A round of funding with nearly $9 million raised from the likes of Google Ventures, Lux Capital and Sequoia. Earth Imaging Journal Editorial Director Matt Ball recently spoke with James (Jimi) Crawford, founder and CEO, about the company’s processing power and its ability to provide actionable information to a diverse group of industries.
Crawford: It has been in the works for a while. As many startups do, you start raising your A round as soon as you close your seed round. In that sense, we’ve been working on it for a year, but the serious work has occurred in the last couple of months, getting the people lined up and figuring out who is going to contribute how much to the round.
Crawford: The focus of my Ph.D. program was artificial intelligence (AI). At NASA I ran the Ames work on autonomy and robotics, including plans for Mars rovers, genetic algorithms for optimizing spacecraft antenna design and scheduling algorithms for optimizing Earth observation satellites. We worked broadly across NASA, supplying AI innovation in many areas.
Soon after leaving NASA I started a Google Books project in which we scanned many books. You start with a book, take an image of it, and then at the end of the day if someone types “to be or not to be,” they get a page from Hamlet. There we used AI techniques in the cloud to scale up to the exabyte level for processing. I’m excited to bring that approach to the space industry, exploring how we can automate and do things at a scale that wasn’t possible before.
The thing that motivated me most to get into Earth observation happened when I was working for a company that worked with Skybox. I was in the company’s clean room, looking at the new Skybox satellites being built. I understood that soon Skybox, or some other company, would be flying hundreds of satellites, and the whole pipeline for satellite interpretation would have to change. In a real sense, that’s what inspired me to start Orbital Insight.
Crawford: My lead engineer and myself have been in the data science industry for a long time, and we were talking the other day about how much easier it is to do what we’re doing thanks to cloud computing. We’ve worked on projects where we’ve used more than 100 computers on the Amazon Cloud. It would be incredibly complicated if we had to set that up somehow in a computer center in our office.
Also, on the software side, we heavily use open-source software like GDAL, QGIS, PostGIS and a host of related technologies that constitute our software stack to process satellite images and perform the data science.
Crawford: It’s essential for what we’re doing. I don’t know about the thick-client world. Everything we do happens in the cloud and has to be scalable. When we write our algorithms, we assume we’ll be doing runs of a million images, and we have done runs on more than 1 million parking lot images for car counting. Everything has to run in the cloud; it has to be completely thin client, and in many cases no client—just part of a Puppet workflow that gets substantiated on a bunch of servers in the cloud.
Most traditional software is designed to run on a desktop machine and have a user perform image analysis. That works great when images cost thousands of dollars each and a company has time to spend weeks staring at them. But with smallsat and unmanned aircraft system (UAS) technology, the future is more about processing thousands—even millions—of images at a time in the cloud. That forces us to the open-source model, where we don’t have to pay license fees and the software is designed to run as part of an automated workflow at scale.
Crawford: That’s right. For example, we’ve used convolutional neural networks to count cars. We had humans mark a few hundred parking lots, taking a mouse and putting a red dot on all the cars. That provides the triggering mechanism for the neural network, and then the neural network learns what a car is. Then we processed a million parking lots through the neural networks and counted millions and millions of cars in those images.
That’s the kind of automation at scale that lets you see the world in new ways. That takes you from looking at one parking lot to answering big-picture questions such as “How has the United States recovered from the recession in the last five years?” or “How did it recover in one state versus another?” or “How did Target do versus Walmart?” In other applications we’re looking at how China is growing, how many new building are going up in cities across China and how different regions are doing. It’s all about automating the algorithms so we can process thousands and thousands of images in parallel and get a sense of what’s happening on Earth on a macroscopic scale.
Crawford: Neural networks are based on and inspired by the shape of actual neural networks in the brain, but they become a mathematical construct that’s generalized a bit from the original biological inspiration. Basically, a neural network uses a higher-order statistical analysis based on a set of algorithms that take an input, such as a bunch of pixels that compose an image of a parking lot, and an output that looks at a classification—for example, “car” or “not car.” It learns the association among certain patterns of pixels.
It’s a nice technique for us, because if we later needed to train a neural network to recognize a tractor, the top of a building or a hospital, then we could mark those objects and use much of the same code to learn the pattern of cars versus tractors or buildings versus hospitals.
Crawford: You can, but the resolution of the imagery available today varies significantly. If you use DigitalGlobe’s WorldView-3 images, you have 30-centimeter accuracy and really nice optics. With a WorldView-3 image you really could tell the difference between a car, a van or a truck. If you use imagery from an older DigitalGlobe satellite, say IKONOS, you’re dealing with a 70-centimeter image, so you’ll be limited to determining whether it’s a car or it’s not.
As we move into the future and start using drone imagery that might get down to just a few centimeters in resolution, then we could literally classify cars by make and model. But that level of resolution is still well into the future.
Crawford: Absolutely! We’re interested in answering questions that make sense to somebody who has never heard of and knows nothing about satellite imagery. We predicted that Home Depot was going to have a blow-out quarter by looking at car counts during the last five years. We built a regression model for those 20 quarters that showed the relationship between the car counts and the company’s reported sales. That way, halfway through the current quarter, we could start predicting what the sales are going to be.
We can do the same thing to predict U.S. corn yields, construction rates in China and all kinds of other applications. Getting to these answers starts with the same pattern. We start with detailed technical work, using machine vision to be able to count something important, and then we lift that up by running it at a national or global scale. We can correlate that information, by analyzing data from the past, to arrive at something decision makers care about now, whether the decision maker works for a government, an investment firm or a corporation.
Crawford: The DigitalGlobe team has helped with that. The company has flown high-resolution satellites for more than five years, so we can use that archive of historical imagery to do really nice correlations between satellite-derived car counts and retail sales by retail chain. Similarly, regarding crop yields, we can compare more than a decade of Landsat imagery with known corn yields we can get from the U.S. Department of Agriculture.
Crawford: One of the really interesting applications in that realm is tracking and measuring deforestation at national scales. It’s difficult to access a lot of heavily forested areas in countries like Brazil or Malaysia. With a satellite image, we can really tell how big the trees are and how deep the forest canopy is. We can also get a good idea of how much deforestation is occurring. That’s definitely an application area that’s on our road map, as it’s a great use for our technology.
Crawford: It’s absolutely a motivator.
I was responsible for Climate’s engineering and data science team through the Monsanto acquisition, but Climate has a different focus. It’s primarily focused on providing insurance and advice to farmers, providing insurance for crops as well as giving advice at a sub-field level—for example, how much nitrogen to apply, how much water to use, when to water, which seeds to plant in different parts of the field, etc.
Our companies are similar in that they work with big data, but the end application is different. We’re focused on providing advice to governments, corporate officials and investors to help them make decisions that impact the world at scale and understand things in a macroscopic way. However, a lot of the technical and big data inspirations are similar.
Crawford: We sell our data as a service—essentially software as a service, but we sell data. A lot of our customers are in the hedge-fund space, because we made a seed round that cared most about the time to close a contract. The great thing about selling to hedge funds is that we have closed contracts in as little as two days from the time we made our first presentation. If you sell to the government, that timeframe can be more than a year.
So in the hedge-fund space we’ve sold a quarterly subscription for data as a service, and the companies that have paid the subscription fee can login to a proprietary website to receive a set of charts and graphs that quickly communicate how each of the retail chains is doing, how the corn crop is progressing, etc. Companies can also download the data behind our charts and graphs to do their own analysis.
One of the major goals of the A round is to move outside of the hedge-fund space and increase our subscription market in government agencies, Fortune 500 companies and a variety of vertical markets. We’re planning to keep the same go-to-market strategy, with data as a service and an appropriate set of charts and graphs to help decision makers quickly understand what the data show. We also have nonprofit partnerships, including one recently announced with the World Resources Institute that’s designed to create models to predict deforestation and monitor the global supply chain for certain commodities like palm oil.
Crawford: We have partnerships with DigitalGlobe and Airbus to secure current, high-resolution satellite imagery. We’ve had detailed discussions with all of the next-generation smallsat providers about partnering with us when they begin producing imagery at scale. We’ve also had discussions with a half-dozen UAS companies, and it’s kind of the same situation. It’s early, but we expect to pull data in from them when they start having data at scale. Our preferred and general model is a revenue-share model where we pull in imagery we need for our data product and share a percentage of the revenue we make with the imagery provider.
Crawford: We’re watching that space closely, and there are a lot of business models. Obviously, the low-altitude UAS market is interesting because the resolution is fantastic. There’s also another UAS market developing at 60,000 feet, and that’s interesting because it essentially entails developing hovering, solar-powered devices that can stay up for a month at a time. You can do some fascinating things with that kind of platform.
Crawford: We call this a macroscope. The microscope was created by biologists to see a cell and other small things. What we’re building lets you see things that are too big for the human eye to see. Our approach lets you look at entire countries or regions all at once, which may encompass a trillion pixels. It definitely allows for a new level of transparency and a new way of seeing Earth systems at scales that weren’t possible before.
Crawford: In addition to continuing to build our core business across new verticals, we’re excited to work with nonprofits to create a new source of data and transparency that will help combat climate change as well as hold governments and businesses accountable for their decisions. For example, we’re working with the World Resources Institute’s Global Forest Watch initiative to create predictive models that may stop deforestation before it starts.