In Season Spotter Image Marking, one of the tasks we ask you to do is to outline individual trees and tell us if they are broadleaf trees or needle-leaf ones.
It’s easier to see different trees at different times of year. For example, evergreens are easier to see in winter when nearby deciduous trees have lost their leaves. Likewise, deciduous trees whose leaves change color are easier to see in autumn. And different lighting conditions change the highlights and shadows around different trees, making some easier to see in one image and others easier to see in other images. So we’ve taken a handful of images from each site in each year, and we show these pictures to you.
Then we combine your classifications across images. Remember that we only ask you to outline three trees per image. But not everyone outlines the same three trees. In the end we get a good sampling of all the most easily delineated trees in the image.
Here is the view from the Arbutus Lake PhenoCam in Huntington Forest, New York:
And here’s what it looks like if we put everyone’s markings for this site on top of one another:
Here, white shapes are broadleaf tree markings and yellow shapes are needle-leaf ones. As humans, it’s pretty easy to pick out the major trees. We also see a lot of stray shapes. For example, sometimes people outline entire vegetated regions instead of single trees. This is probably because another question in Season Spotter Image Marking asks volunteers to do just that, and it gets a bit confusing.
What we really want is a single shape for each major tree. So I implemented a clustering algorithm that takes all these shapes and finds those that are most similar to one another. It groups all these shapes into clusters, and hopefully each cluster represents one tree. What’s tricky is that we have to define a threshold of just how similar the shapes must be to be in the same cluster. If we require that the shapes be super similar, then shapes around the same tree that are just slightly different don’t get grouped together. But if we are too lax about our clustering rules, then multiple nearby trees — especially small ones — get grouped together in a single cluster.
Once we have our clusters, we need to take all the shapes in that cluster and combine them into a single shape. I did this by taking a few of the smallest shapes in each cluster and taking their geometric union. This gives a conservative estimate of the extent of the tree crown. I did it this way because the next step is to run some automated algorithms on the resulting regions, and I wanted to make sure I wasn’t including any pixels that were outside the focal tree.
Here’s what the clusters and resulting “consensus shapes” look like:
The next thing to do is to run our automated algorithms on each outlined tree and compare that with the results when we run it on the whole landscape. Traditional phenology measures have been on-the-ground and done on a tree-by-tree or plant-by-plant basis. And satellite phenology measures integrate across broad areas. This analysis will help link those two types of phenology measurements, so that we can scale-up from the ground observations and understand the satellite observations more biologically.