Original article was published by Taylor McDowell on Artificial Intelligence on Medium
Viewing the trees at such scales can be pretty eye-opening to the effects of natural and human-caused phenomena on the landscape. The river bottoms of the American Southeast are such an example, where rich alluvial soils are a boon for agriculture. The conspicuously deforested lands around the lower Mississippi River delineate the historical floodplain nearly spot on. However, tree canopy data goes beyond illustrating such well-understood quirks about our landscape and land-use practices. Accurately mapping trees at this scale also can be used to monitor trends in forest cover, determine where trees could be replanted to bolster erosion control, or to identify candidate properties for conservation easements.
How accurate is it?
Great question! We estimate our nationwide tree canopy to be 96.6% accurate (hey urban planners! That figure is 97.3% for TIGER designated urban areas). We randomly sampled approximately 48,000 points, and with the help of very dedicated staff, we manually decided which points landed on trees, and which did not. This way we can evaluate how well our CNN model predicted trees from the imagery by comparing the results to our own predictions.
Putting data to work
While these trees are sure nice to look at, this data is meant to work! Whether you are a data scientist, geographer, urban planner, natural resource manager, insurance analyst or cartographer — however trees or forests matter to your work, accurate and reliable data is now available. The US Tree Map accurately maps tree canopy data with 1-meter resolution for the continental United States, making it the largest archive of high-resolution tree canopy data available.
For us at EarthDefine, this isn’t the end of the road. While we are pretty happy about our new tree canopy data, we are already working to improve it. Trees and forests change in many ways and take on a multitude of appearances in aerial imagery. We want to ensure that we can accurately map all such changes— things like bark beetle infestations, wildfires, regenerating trees from recent disturbances, etc. The same goes for the variety of NAIP image-specific variations, like shadows in mountainous regions, leaf-off deciduous trees, or fall foliage. Our goal continues to be to provide the most reliable and accurate data, so end-users can do their work with confidence.
Over the next few weeks, we’ll explore how these data can be leveraged for scalable solutions and analyses in a series of briefs. We’ll dive into as many different applications as we can to show the benefits of good data. For those interested in exploring the data themselves, downloadable samples are available on our website. Stay tuned.