The Center for Data Innovation spoke with Dr. James Crawford, chief executive officer of Orbital Insight, a satellite data analytics startup based in Palo Alto. Crawford discussed how his machine learning experience from Google Books led him to work with satellite data, and how analyzing images of parking lots can lead to powerful economic insights.
Joshua New: Orbital Insight specializes in using machine learning to analyze satellite imagery. Why is machine learning well suited to this task?
James Crawford: When the first satellite images were taken, each one represented a massive investment. Highly trained scientists looked at each individual frame to find insights from those early images. Now, satellite images are gathered every day from dozens of satellites in the sky. In a matter of years we’ll have daily pictures of every corner of the world. In a few decades, satellite networks will be so extensive that we’ll be able to get near real-time images of anywhere in the world. Having people analyze each image is extremely difficult today and will soon become impossible unless we want to draft an army the size of the population of New York City.
At the same time, machine learning and big data processing are coming of age. Cloud computing gives us the ability to process a nearly unlimited number of images, and techniques like artificial neural networks can detect subtle patterns and changes in images. We can count and recognize cars, trucks, trains, shipping containers, new home construction, health of corn fields, and skyscrapers under construction. And once we can analyze one image we can quickly scale to analyzing images from every country in the world.
New: You use data provided for free from the U.S. Geological Survey’s Landsat 7 and 8 satellites, so your customers are paying for your analysis, rather than the raw data itself. What other data sources do you use that your customers may not be able to otherwise access?
Crawford: We work with all of the major satellite image providers: Planet Labs, Digital Globe, and Airbus being the bigger private sources. Our analysis is two steps removed from the raw data: we count or measure objects across a country or even the world, and then we aggregate that information into meaningful insights. We aggregate data into trends, provide context, and compare it to expectations. This lets our customers, who are faced with making decisions with incomplete data, have access to an unbiased source of real-world numbers.
New: You previously worked as the engineering director of Google Books. What is the relationship between your work there and satellite data analytics?
Crawford: At Google Books, we wanted to be able to type “to be or not to be” into Google and find the full text of Hamlet. But first, we had to digitize millions of physical books and turn those pages into computer-readable text. There we were picking out letters from images, and now at Orbital Insight we’re picking out cars, lakes, and buildings. The type of question that we want to ask now is, “how many cars parked at the Super Bowl during the game?” and get a number back. With both projects, the input is a huge quantity of images that we use computer vision to analyze. The difference is what we’re pulling out of those images.
New: Can you discuss how you were able to predict retail store revenues after your algorithms learned to recognize cars?
Crawford: Our algorithms analyze car counts at 50,000 retail chain parking lots. Of course, no human team would be able to count that many cars. The main assumption that we had to confirm was that people park close to the shop that they’re going to. We did some case studies and found that to be the case—if someone’s car is parked near a store in a mall, that is the store they are going to. Once we know where the lots are, we run the algorithm over those locations repeatedly. Then we leave it to the data scientists to extract the trends, and control for the time of day and other variables.
By counting cars in retail chain parking lots we’re able to measure, over time, whether a store’s sales are accelerating or declining. We’re not the only people who are making these guesses, but we get it right over 70 percent of the time over the last few years. We also release our prediction a few days before the official numbers come out, so our customers know what to expect ahead of time.
New: What kind of insights do you hope to be able to reveal as your technology gets better? Anything unexpected?
Crawford: Our goal at Orbital Insight is to unite the data of the global economy and supply chain. When we’re watching for deforestation in Brazil, we’re tapping into the palm oil supply chain. When we look at satellite images from China, we can take measurements on a market that doesn’t share reliable data with the world. We have a long list of applications, each taking us deeper into the hyperlocal indicators that we can aggregate into worldwide insight. One of our newest signals, a surface water detection algorithm, can give advanced warning of water shortages wherever we look. That kind of information, when it gets to the right people, will help stabilize economies across the world.