Since 2014, 49,520 registered volunteers have classified penguins, chicks, and eggs in Antarctic photos; more than 6 million classifications helped build a colony dataset

Somewhere in Antarctica, a camera snaps another picture of a crowded penguin colony. Far away, a person at home clicks on every bird in it. In 2018, Oxford researchers published a study in Scientific Data showing that more than six million images had been classified on a project called Penguin Watch by 49,520 registered volunteers and many anonymous participants. This created a dataset scientists could compare with expert counts.

Why counting penguins in Antarctica is so hard

Antarctica is remote and harsh. Large-scale monitoring on the ground is difficult and therefore rare, the paper points out. As a result, many findings come from a few locations and are then generalized to other areas. However, this approach can end up missing important differences since threats such as overfishing often vary from region to region.


Cameras offer a way around this. They are placed above penguin colonies and take pictures on a regular basis, generally once per hour year-round. By 2018, more than 150 cameras were watching penguin colonies in Antarctica and on some Southern Ocean islands, and the study's data came from 14 of them. That is good for science, but it creates a new problem. All those photos are hard for research teams to process manually.

How ordinary people did the counting

The Penguin Watch project began in September 2014 on Zooniverse, an online crowdsourcing platform for scientific research. The task was simple. Volunteers saw one photo, clicked on all the penguins, and then classified them as adults, chicks, or eggs. There was a category called 'other' which was used to classify humans, ships, and other animals. When visibility was poor, for example, in low light or with a blocked lens, volunteers could answer "I can't tell" to the question of whether any animals were present.

A person could miss a bird, so repetition was necessary. For example, if any volunteer spotted an animal in a photo, the image was shown to ten people by default, and software grouped their clicks. When several of them clicked on the same bird, the software merged those clicks into one "consensus click," representing one animal. It worked somewhat like a vote on where each penguin was standing.

How well did the crowd do?

The counts were validated by comparing volunteer and expert counts at four camera sites. For the adult penguins, the best results were when at least four volunteers clicked on the same spot. At that threshold, the average difference between crowd and expert counts was 0.9 to 2.4 penguins per photo, depending on the site. In a separate test, an expert had marked 100 photos as empty, and volunteers agreed on 96 of them. Three of the four disagreements were stray clicks from a single volunteer, and the fourth photo held animals the expert had missed.

However, the technique had flaws, and the team flagged them. Chicks would often go unnoticed, possibly because they hid behind their parents. The busiest colony averaged about 36 penguins per photograph and produced the most errors. They concluded that, with careful filtering, Penguin Watch can analyze large numbers of images.

Where computers fit in

A 2020 follow-up study from largely the same team compared two ways of getting the counts: volunteers and a computer program called Pengbot. Pengbot was first described in a 2016 conference paper by Arteta, Lempitsky, and Zisserman. It was an early solution for counting from crowdsourced dot clicks, and volunteer clicks were used to train it.

The results were mixed. The program's counts were, on average, less than two penguins off an expert's count at three of the four sites. In the case of the most crowded spot, the difference was around 12 penguins. The computer struggled with overlapping birds and penguins at the edge of the frame, but humans handled them more easily. The team called the two methods complementary.

What a crowd can teach us

Penguins are among the most threatened seabirds, with threats including overfishing, pollution, and climate change, the 2020 paper noted. Steady counts from multiple colonies can allow scientists to understand how their numbers shift. Comparing counts across later years, the authors added, may reveal population trends that can help conservation efforts. The images, clicks from volunteers, and all the additional data are publicly available for use.

There is a human side, too. A 2014 Zooniverse survey, cited in the 2018 paper, found that 90.6% of the respondents reported that they enjoyed helping to advance science. These were volunteers from the entire Zooniverse community, not just penguin lovers. However, it shows why projects like this can work. There are people who want to help, and Penguin Watch just gave them a small, clear job that added up to something useful.