Sat Summit is the semi-annual conference held in Washington D.C. that brings together leaders in the earth observation, international development, geospatial analytics, and open source software industries. Every year is amazing, and this year was no different, thanks mostly to the talented and hardworking event organizers at Development Seed and Mapbox. One of the many reasons I’m so smitten with this conference is the culture of openness and honesty that has become increasingly rare in the self-absorbed world of “high technology” I frequently find myself in. If you’re like me, and you’re more interested in understanding the real state of the earth observation industry rather than being sold stuff, Sat Summit is worth your while.
Here are a few big takeaways from my trip to Sat Summit:
1. Machine learning isn’t revolutionizing the earth observation industry… it’s creating an entirely new industry.
The hype of “new space” and the value of cheaper, timelier, and higher resolution data was premised on a flimsy assumption — that more data would beget more information. Instead, our industry has suffered from a runaway ratio of noise-to-signal. I once saw a presentation from the National Geospatial Intelligence Agency (NGA), the world’s premier organization for analyzing geospatial imagery at scale, which called the phenomenon of having too much imagery and nothing to do with (most of) it a “success-catastrophe.”
Machine learning, and deep learning employed for image recognition in particular, ruled the day at Sat Summit. The opening keynote focused on deep learning applications, every sponsor but one had a demo or showcase project that leveraged deep learning, and every interval of panels had a topic that directly or indirectly touched on applied deep learning. Maybe someday I’ll write a whole blog post about this…but my conviction that earth imagery is but a small subset of the much larger and more transformative industry of image recognition is now stronger than ever.
How all-consuming is this shift to focusing on machine learning? Let’s take the lightning talks from the “Environment” track:
— Yoni Nachmany (he/him) (@YoniNachmany) September 19, 2018
- Astraea, on how they use machine learning to track deforestation in the Amazon
- Jupiter Intelligence, on how they use SAR data and machine learning to predict flood risk
- Development Seed, on how they work with NASA to predict hurricane strength more quickly than traditional methods with deep learning
- Azavea, on how we use deep learning on aerial imagery to catch pipeline disasters before they happen with our partners American Aerospace
If you’re keeping score at home that’s 4/4 on a random track not explicitly related to machine learning or computer vision…
Even just a couple years ago, demonstrating the ability to do real-time search and visualization of a massive satellite imagery archive, like Landsat, was considered a technological breakthrough — I wrote a whole blog post about it. At the last Sat Summit, Bronwyn Agrios gave the most memorable keynote of the day about Astro Digital’s efforts to build such a platform. But at this years’ Sat Summit, the conversation had noticeably shifted from how do we make access easier to how do we make access unnecessary? In other words, we need to stop building beautifully engineered tools for an imaginary audience and start building solutions for real people in the field.
Why don’t we need to belabor how to make data more accessible? Because there’s already a well-worn playbook that has been proven to be successful. One of my favorite panels included Kristi Kline, the project manager at USGS tasked with managing their decades-deep Landsat archive. She, like many of her colleagues in government, is one of the unsung heroes of the open data movement, and has shepherded Landsat from an era when the archive was stored on (literal) truckloads of magnetic tape to today, when the whole archive is accessible online. Pair Kristi’s inexhaustible commitment to making open data more accessible with the incredible work of Earth on AWS to provide Landsat 8 for free on their platform, and you get a chart that looks like this:
— Chad Blevins (@geocruizer) September 19, 2018
However, despite the clear success of programs like Landsat and NAIP, there are overt threats to the continued openness of these datasets. NAIP was recently put under the microscope with proposals to take it private and narrowly escaped with budget still intact. Tom Lee, Policy Lead at Mapbox and prolific Twitter luminary, gave an impassioned presentation (slides available here) about the need to stay vigilant in which he implored the audience:
Today and tomorrow you’re gonna hear about a ton of exciting and innovative work — I encourage you to think about the open data it’s built on top of and how we can preserve it.
Thanks again to our friends at Development Seed and Mapbox for hosting an incredible conference, and to the many excellent speakers who shared their work. I personally can’t wait for the next one.