Marking shifts in time-series

One of the tasks we ask you to do in Season Spotter (Image Marking) is to mark places where there is an obvious shift in the camera’s view over time. We do this by extracting the center vertical line of pixels of an image.

We do this for an entire year’s worth of images (one image per day) and then glue them together side-by-side. The result looks like this:

2011

One really nice aspect of this technique is that it becomes visually clear when there is a change in the view of a camera. You can generally see the horizon line and this and other features pop up and down when the view of the camera changes.

In this case we see two major shifts. The first is about half-way across when it appears that either the camera was completely replaced with a new camera or else the settings on the camera were changed. We know this because the size of the image changes. That’s what the black bar on the top right tells us, as we align our center line of pixels at the bottom of the composite image.

The second shift takes place towards the end of the year. We see a slight shift in the horizon and a more major one in the vegetation. The image size remains the same, so it’s possible that the camera was bumped, moved to a new location, or simply adjusted.

If we go look at images from this site, we can see how the camera’s view actually changed:

Early 2011

September and October 2011

December 2011

We can see that the shifts are not major ones — the same general scene appears in each of the three camera views. This is good because we can compensate for these relatively small shifts in our automated processing.

In particular, we create regions of interest (ROIs) by drawing shapes on the images. These ROIs act as masks when we calculate greenness over the year, so that we are using the same general vegetated region for each of the three scenes. In this case, we would draw ROIs to try to get a good representation of the area of close forest that appears in every scene.

When we create separate ROIs for each scene, we’re often able to recover what looks like a smooth greenness line. If we don’t correct for camera shifts, we often get a greenness line that suddenly jumps from low to high values or vice versa.

I should note that this technique of looking at vertical center lines does not catch every camera shift. Sometimes cameras shift only horizontally so that the horizon remains the same and there are no major indications of change in the vegetation. But it does catch the majority of them. By marking the shifts, you help us figure out where we need to adjust our ROIs for automated greenness processing.