- A multinational research team has deployed time-lapse cameras at various penguin breeding colonies to enable a widespread, long-term study of these top predators in the Antarctic ecosystem.
- Volunteers have played a critical role in processing the millions of images resulting from the multi-year study to better understand reproductive behavior and nest success rates across the Antarctic Peninsula, South Shetland Islands, and South Georgia.
- Citizen scientists can help produce large data sets needed to train artificial intelligence algorithms.
Cameras add monitoring power
Field data collection is challenging in the best of conditions, and in an environment as harsh as Antarctica, large-scale, long-term field monitoring studies are rare.
To get around the problem, a multinational collaborative research effort has installed time-lapse cameras at more than 30 penguin breeding colonies in Antarctica and remote islands in the Southern Ocean. The network of cameras takes images of the penguins year-round, enabling researchers to monitor the health of the colonies over time and space by documenting nest survival rates and comparing changes in population dynamics and reproduction to shifts in temperature and human fishing activity. All without people being present.
Unlike motion-triggered camera traps, time-lapse cameras automatically take an image at set intervals, such as every hour, even if no motion or animal is detected. Scientists use them for automated monitoring of field sites, to capture a near-continuous record of occupancy and observations of a population or habitat over time.
In Antarctica, the researchers use wire and stones to hold each stationary camera in place and position it to film multiple nests associated with, in some cases, more than 30 penguins. Once a year, workers on the ground replace the batteries and storage cards in the cameras.
The authors say the images from time-lapse cameras have several advantages over cameras with a passive infrared (PIR) sensor triggered by body heat. Time-lapse cameras can show penguins, even at a far distance, which a passive infrared (PIR) sensor may not detect. Collecting data from constant time intervals enables the researchers to perform analyses, such as mark-recapture studies.
A given camera may lack sufficient power to continue operating throughout an entire year with a PIR sensor. Increased function in the summer months could drain the battery and leave a colony unmonitored during crucial stages of the birds’ annual cycle (e.g. the arrival of adults in the spring).
Similar to camera traps, time-lapse cameras set out over time generate large image data sets, with many thousands or millions of images, the analysis of which costs time and money. The network of 91 time-lapse cameras around Antarctica has generated millions of images, each of which might, or might not, contain penguins.
Processing 6 million images
The effort has enlisted more than 50,000 volunteers over four years to annotate the images through a citizen science program called Penguin Watch, hosted by the Zooniverse. Penguin Watch offers the public access to thousands of images of penguins at colonies around the Southern Ocean and enables people to contribute to conservation research. Collectively, Penguin Watch volunteers have annotated more than 6 million images of gentoo, chinstrap, Adélie and king penguin colonies.
The research team recently described its citizen science program in Nature’s Scientific Data journal. They assessed its success using a sample of nearly 74,000 images from 15 of the cameras in the network. Each camera generally captures one image every 30 or 60 minutes, between 7 a.m. and 8 p.m., year-round, and images are linked to their corresponding date, time, and temperature readings.
The volunteers classify images by tagging individual penguins as “adult,” “chick” or “egg” and other animals, humans or ships as “other.” Several volunteers review each image to increase data reliability, and if a volunteer identifies animals in an image, the platform shows that image to 10 other volunteers. If none of the first four volunteers finds an animal in an image, that image is removed from the active data set.
“Volunteers do not need to identify the penguin species, as most of the colonies we examine are single-species colonies,” lead author Fiona Jones of Oxford University told Mongabay. “If they wish, our volunteers can ‘flag up’ an interesting image on the Penguin Watch talk forums — for example, they might notice some unusual behaviour. Our online moderators will then be able to provide information about what they’ve seen.”
Citizen scientists have proven a valuable resource to help researchers extract specific information from the resulting large image data sets. By categorizing images by their primary subject, including those with no animals, the Penguin Watch volunteers facilitate subsequent analyses that are more in-depth, such as detecting chick hatching or other short-lived breeding stages.
“The response from the general public to our project has been phenomenal — and not something we expected at all,” Jones said. “We simply couldn’t do our research without them.”
A 2014 Zooniverse survey found that 90.6 percent of participating volunteers said they “like to contribute to scientific progress,” while 84.7 percent of people were “fascinated by the projects” in which they were involved. Individual responses also conveyed an enthusiasm for the projects and a desire to discover something new.
“In order to attract volunteers, it is important to let them know why they are helping you,” Jones said. “They are giving up their free time to assist your research, so it’s great to give regular updates on how your project is progressing, and what all their hard work is going towards.”
Where volunteers meet machine learning
Penguin Watch launched in 2014, when volunteers began reviewing images from 2012 and 2013.
Now, scientists are beginning to employ machine learning techniques to automate the image data extraction process. But where do machine learning algorithms get their training data?
The volunteer data is one new source of big data sets. The Antarctic penguin volunteers have generated a large amount of information that can be used to train machine learning algorithms to carry out the task automatically.
In fact, “The data exported from the Penguin Watch project have already been used to train machine learning algorithms,” Jones said. “The ultimate goal is to be able to use citizen science and machine learning in a complementary way, rather than one replacing the other.”
Jones, F.M., Allen, C., Arteta, C., Arthur, J, Black, C., Emmerson, L.M., Freeman, R., Hines, G., Lintott, C.J., Macháčková, Z., Miller, G., Simpson, R., Southwell, C., Torsey, H.R., Zisserman, A., & Hart,T. (2018). Time-lapse imagery and volunteer classifications from the Zooniverse Penguin Watch project. Scientific Data 5(180124).
FEEDBACK: Use this form to send a message to the editor of this post. If you want to post a public comment, you can do that at the bottom of the page.