When we started AI for Earth, we had one simple but huge ambition - to fundamentally transform the way we, as a society, monitor, model and manage Earth's natural resources.
That transformation will ultimately require collecting and processing exceptionally large datasets - an endeavor that can take a lot of time and money, even with advanced cloud computing and AI tools like deep learning. In part, these barriers have curtailed progress on important tools for conservation, like up to date land cover maps.
I'm excited to share that we've made a computing breakthrough that moves the needle towards real-time analysis of land cover data. We first shared the news at Build -Microsoft's annual developer conference-and on Thursday and Friday of this week AI for Earth's principal engineer Jennifer Marsman will be discussing our results in detail at an AI event in Paris.
Why does land cover mapping matter? There are three big reasons.
Land cover mapping is the foundation of effective conservation and sustainable growth. Data is the lifeblood of conservation efforts; and to protect complex ecosystems, such as watersheds, conservationists need accurate and precise spatial data. Real-time, high resolution land cover maps can guide conservation efforts, but creating these maps using available imagery-and tracking changes over time-requires complex algorithms and computing resources.
This foundation has been in shambles. The best available land cover map in the United States is at 30-meter resolution and eight years out of date. That's because processing the explosion of satellite, sensor and aerial images is tedious and time-consuming.
This situation is only going to deteriorate. We are now collecting geospatial data at an incredible rate. We need algorithms, and the hardware they run on, to be able to keep pace with the increasing speed of data collection.
Because this problem of up-to-date land cover mapping is so basic and so important, it was one of the very first projects we took on with AI for Earth, in partnership with Esri and the Chesapeake Conservancy. Using algorithms on Microsoft's Azure platform and integrating with Esri's ArcGIS spatial mapping software, the Chesapeake Conservancy and its collaborators in the Chesapeake Bay Partnership created an accurate, current land cover map of the Chesapeake Bay watershed at one-meter resolution-giving conservationists access to data with 900 times the information that was available before.
That's great for the Chesapeake, but it still left the rest of the country to be mapped, a task that would require processing over 10 trillion pixels of imagery into categories like forests, fields, water, and urban areas. Until today, this would take a huge amount of time and manual resources.
Now, through Project Brainwave, we are capable of processing more than 20 terabytes of aerial imagery into land cover data for the entire United States in much less time, and for much less money, than existing solutions. We are using a new FPGA (field programmable gate array) chip solution in Azure, which can plow through nearly 200 million images in just over 10 minutes for a cost of $42. These results pave the way for organizations to produce new, high resolution land cover maps on infrastructure that can scale up or down for all sorts of problems around the world.
To be clear, algorithms need to be both fast and accurate, and there's still a lot of work and testing to do on that front. Nonetheless, these speedy results are a good first step in empowering people to apply AI at earth scale. And, of course, land cover mapping is just one of over 100 projects in which we have invested - please check out our website: https://www.microsoft.com/en-us/aiforearth for the latest updates on our grantees, projects and progress.
To learn more:
Tags: AI for Earth, Azure, Environmental Sustainability, FPGA, Microsoft, Project Brainwave, Sustainability