Last week I described some initial results from the Spring Challenge. I showed how we used individual classifications to build a dataset for a single site in a a single year. And we discovered that using paired images 7 days apart and then smoothing the classifications gave us a good estimate of the “start of spring” date and “end of spring” date for that site in that year.
Since then, I’ve been comparing these “start of spring” and “end of spring” estimates with other estimates that we get that are automatically derived from our greenness time series. To get these automated estimates, we first take the daily greenness values and draw a smooth curve through them. Then we look at the total amplitude (height) of the greenness signal and pick the date where the smoothed curve passes through 20% of that amplitude. We call that the “start of spring.” We call the date where the smoothed curve passes through 80% of that amplitude the “end of spring.”
We can plot our Season Spotter estimates on a curve, along with the automated estimates to get an idea of how the two compare. Here is a plot from 2014 using data from the “canadaOA” camera at Prince Albert National Park, Saskatchewan, Canada — the same site and year we looked at last time.
Each of the green dots represents the greenness measure for a single day. The black line is the smoothed line that is fit through the green points, and the gray region around it is our certainty range. The two orange squares are the automated estimates of “start of spring” and “end of spring.” The blue squares are the estimates of “start of spring” and “end of spring” from Season Spotter. Both the orange and blue squares may have horizontal lines coming out of them showing the likely range of dates for these estimates. The longer the line, the less confidence we have in exactly where our square lies. For all the squares, I’ve drawn dotted lines from them to the smoothed line so we can visually compare them more easily.
As we suspected from our analysis last time, the estimates from pairs of images 1 day apart and 3 days apart are closer to the middle of spring, where it’s easier to see the change in leaves. The estimates from the 7-day apart images look very good — even better than the automated estimates!
If we look at “start of spring” estimates from 7-day apart images from all the sites and years that we put into Season Spotter, we see a trend:
Here, each site has a different color and each rectangles is a year. So, for example, there are three purple rectangles showing three different years from the canadaOA camera. The lines coming out of the rectangles show us our certainty, as before. Going across is the “start of spring” from the automated method. And going up is the “start of spring” from Season Spotter classifications. The diagonal dotted line is the one-to-one line. If all our rectangles were on this line or scattered evenly around it, it would mean that the estimates from the automated method and from Season Spotter pretty much agree. Instead, we can see that Season Spotter regularly predicts an earlier spring than the automated method, because most of the rectangles lie below the one-to-one line. This suggests that we might want to tweak our automated process to use a lower threshold.