September 8, 2016
You may have noticed that certain plants are flowering earlier than they used to just a decade ago – or that trees are often putting on spring leaves earlier. The timing of events in the life of plants is known as phenology. Plant phenology is very sensitive to weather patterns, and as climate changes, so too does the timing of flowering, seed production, leaf-out, and other life stages.
Because of its sensitivity, plant phenology is something scientists study to understand how climate change is affecting life on Earth. We use elevated digital cameras, called PhenoCams, to track the greenness of landscapes all over North America. These cameras take pictures every half-hour, all year long. When we analyze them, we can see when leaves first come out, when leaves are at their peak, and when they turn color or die.
To measure the leaves in each image, we use an algorithm that calculates the image’s “greenness” as a number. This number changes from image to image and over the seasons. We can plot these numbers over time on a graph, giving us the heartbeat of a landscape.
We thought we could get more information out of these images than just the greenness, though. Things like flowers and cones, weather events, and the phenology of individual trees are all visible in the images. The problem is that we don’t have algorithms to extract this information automatically. And we have hundreds of thousands of images, so we can’t look through them all with just the few people in our group.
So, to see if it is possible to extract additional information from the PhenoCam images, we created the online citizen science project Season Spotter. In Season Spotter, anyone could volunteer to help us out by looking at images and answering questions. Questions including things like whether there were flowers, whether there was snow, or which of a pair of images had more green leaves. Volunteers also drew outlines around individual trees.
The questions we asked volunteers reflected three different types of information we wanted to extract from the images:
- Fine details of plant life history, such as flowering and seeding
- Differentiating individual trees from a landscape
- Transitions between seasons
Approximately 50,000 images were analyzed by about 18,000 volunteers, with multiple volunteers classifying each image. What we found was that volunteers could, in fact, extract most types of information. In particular, people were able to identify flowers, cones, and crop stages (with 92-98% accuracy), but they were not able to regularly identify grass seedheads. (Grass seedheads might be too small in the images or too unfamiliar for the average person to classify them well.) Having information about flowers, cones, and crops will be useful in understanding how changing climate is affecting plants more than just when they have leaves out.
Volunteers could also reliably pick out trees from a forest canopy and draw lines around them. We found that using the volunteer-drawn outlines, we could calculate the greenness of individual trees over time. This will be very useful in bringing together different data sources – such as ground measurements and satellite data – to understand impacts of climate change on phenology at different scales.
And volunteer data on the timing of leaf emergence, autumn leaf color change, peak autumn leaf color, and leaf fall provides an unparalleled opportunity to evaluate our existing algorithms for determining seasonal transitions. We found that our algorithms for estimating the start and end of spring are quite good and match the volunteer data pretty well. Our algorithm estimates for the start of autumn does not work well across all sites. And our estimates for end of autumn work pretty well at most sites, but not all. Additionally, the information volunteers provided about changing leaf color versus leaf fall tells us more about what is going on biologically than just a single measure for the end of autumn. All this information allows us to create new algorithms to identify these processes separately in future images.
Going forward, we will be refining the classification tasks in Season Spotter to be more efficient and less repetitive. Now that we know which data volunteers can help us extract from the PhenoCam images, we will pinpoint some specific scientific questions that can be answered with it. In the end, we will know more about how climate is affecting plant phenology in North America.
This is a plain language summary of the paper:
Kosmala, M., A. Crall, R. Cheng, K. Hufkens, S. Henderson, A.D. Richardson. (2016) Season Spotter: Using citizen science to validate and scale plant phenology from near-surface remote sensing. Remote Sensing, 8(9): 726. doi: 10.3390/rs8090726.