What is Raster Vision and why should I care?
The amount of readily available satellite, aerial, and drone imagery has exploded over the past ten years and shows no sign of slowing down as a massive influx of venture capital has propelled the earth observation industry into the modern era of ubiquitous monitoring. It has already been almost three years since Planet accomplished its famous Mission One: to image the entire landmass of the Earth once a day. New commercial sources of imagery like Capella Space’s Synthetic Aperture Radar are coming online this year, promising to provide hourly global coverage. Further into the future, ambitious companies like Theia Group aim to image the entire earth multiple times per second while incumbent Maxar plans to launch a new high-resolution constellation of satellites in 2021 capable of revisiting the same area on earth 15 times in a single day at just 29cm resolution.
In theory (emphasis on theory), every global industry from logistics and supply chain management, to forestry and agriculture, to mining and energy, to public health and humanitarian aid should benefit from this wave of timely, objective data. But while there is tremendous capacity for imaging the earth, there remains tantalizingly scant capacity to actually…do much with those images. The best shot we have is deep learning, the same technology powering autonomous vehicles and AR mustaches in Snapchat. Remarkably, the deep learning revolution that has transformed the field of computer vision has been overwhelmingly accomplished in the open, with core developer tools like TensorFlow (Google) and PyTorch (Facebook) remaining by-and-large free for anyone to use.
Raster Vision is the interface between the fields of earth observation and deep learning, making it easier to apply novel computer vision techniques to geospatial imagery of all types. We built it to solve our own problems working on client projects where we found ourselves constantly reinventing the wheel and taking weeks or months to train and deploy effective models only to have to start over from scratch the next time. It’s an open source, openly licensed, free-to-use developer tool that has revolutionized the way we’re able to work with satellite, aerial, and drone imagery. If you’re an organization looking to take advantage of this same confluence of technological waves, we think Raster Vision may be worth a look.
One of the hardest aspects of building capacity for data science and machine learning across a large organization is simply a matter of communication. If one researcher in a corner of the company comes up with a breathtaking and valuable new analytical technique, how does she share it with colleagues across the company? How will they even understand what they’re looking at or how they might borrow elements of it for their own benefit? The ugly truth of most data science is that underneath a thin veneer of beautiful charts and graphs is a chaotic undercarriage of bespoke, undocumented, impenetrable software meant to work in an extremely narrow context.
One of the major advantages of using Raster Vision is that it simplifies and standardizes the messy parts of working with earth observation data without sacrificing the flexibility required for truly groundbreaking, innovative research. If your team can spend less of their time wrangling and reformatting data, and more of their time experimenting with novel approaches to analyzing that data, overall productivity goes up. But perhaps more crucially, the ease of communicating what has been accomplished in a particular experiment, and making that experiment easily repeatable, extendable, and interpretable is an often underappreciated value that Raster Vision has enabled at Azavea.
Real, practical machine learning is an iterative process that often begins with a question and ends with ten more unexpected questions–the capacity to quickly hypothesize, test, react, and reorient will determine the productivity of a data science team. The more time that is spent on custom, one-off scripting just to prepare an experiment, the less time your team has for the valuable work of experimentation. Additionally, true research and development is uncertain–a certain percentage of projects will fail and be abandoned, and the sooner you can arrive at a decision not to continue investing in a particular avenue of research, the better. Raster Vision won’t help you successfully complete a higher percentage of your research projects, but it may help you complete more projects over the same fixed window of time as it has for us.
Lastly, when considering your strategy for incorporating machine learning into your business or organization, think about who the value of the intellectual property you are funding is accruing to. If you’re working with a private company or a proprietary solution, who owns the rights to the models you are funding the creation of? Ask yourself: is it worth it to me to fund a project that can be resold to my competitors or taken away from me at any time? If the project is not critical to your business or is just part of a transient project, the answer is probably, “yes.” But if what you’re working on is something you anticipate using every day or with every customer, the equation may tip in favor of retaining the intellectual property that will be created as a result of your investment.
As mentioned above, the hammer and nails of deep learning are almost entirely open source and openly licensed, just like Raster Vision. The challenge of adopting these tools and forming them into a system that meets your requirements is still significant, but not so significant that you should cede ownership of that system to a vendor in most cases.