Researchers Use GoogLeNet & Drones to Help in Identifying 1.5 Million Penguins on Danger Island.



A recent study from Stony Brook University’s Department of Ecology and Evolution [1] found that the Danger Islands off the Antarctic Peninsula is home to 751,527 pairs of Adélie penguins.

My first thoughts were- “That’s a lot of Penguins,” followed by “How do a group of researchers go about counting that many penguins? Do they just show up on Antarctica and start counting?

It turns out, the researchers in the study had three methods for surveying the population:

  1. Manually counting nests – So I was right, in a way. The researchers used good, ole mechanical turk to count some penguins. According to the study, this was the best method to identify chicks and eggs.
  2. Counting individual nests from panoramic ground photos.
  3. Using a deep neural network to count individual penguins from unmanned aerial vehicles and satellites.

Naturally, the third method  intrigued me. Not only are they sending a type of drone to survey Antarctic islands (which is pretty cool in itself), but the implementation of neural networks for counting species is something that I had seen before in a Kaggle competition. several months ago. Now, here it is in use in an academic study.

On GoogLeNet and CNN’s

The model that the researchers used is called DetectNet – A GoogLeNet (AKA Inception) based  image detection model[2]. GoogLeNet, like other convolutional neural networks, is used for finding connections in image data. The convolutional layers work by partitioning an image into chunks to account for placement variability (For example, an image with a cat in the corner and another with a cat in the center would both have characteristics of cats, just in different places in the photo). The network processes the images pixel by pixel – to include RGB value and placement – looking for connections between similarly labeled photos.

GoogLeNet architecture from “Going Deeper with Convolutions”

The model placed first in the 2014 ImageNet Large Scale Visual Image Recognition Competition with a top-5 error rate of 6.67% and remains one of the most popular per-trained models for image detection.

The researches reported a 0.6% average difference between automated nest counts and an in situ ground count. This reaffirms, to me at least, the capability of machine learning in species counts. There is value not only in its ability to reduce manual work and speed up reporting, but also to survey areas that are difficult or dangerous for manual counting.

Ecologic Implications

The 1.5 million penguins counted in this study is a whopping three times that of previous estimates [3]. This finding has a couple of implications brought up in the discussion of the article:

  1. Krill predation is probably higher than previously estimated – pretty intuitive, knowing that penguins eat krill.
  2. The Danger Islands have been relatively unaffected by warming, indicating a West-East warming pattern. Naturally, I would have thought that warming went North-South, so this was interesting.

Overall, the discovery of such a large colony of penguins should urge others to consider the Danger Islands for further environmental protection.


  1. Borowicz A, et al. Multi-modal survey of Adélie penguin mega-colonies reveals the Danger Islands as a seabird hotspot. Nature Scientific Reports, Article 3926. March 02, 2018. DOI:10.1038/s41598-018-22313-w.
  2. Szegedy, C. et al. Going deeper with convolutions. Proceedings of the IEEE conference on computer vision and pattern recognition. 2015.
  3. Lynch, HJ. First global survey of Adélie penguin populations. The Auk 131, 457–466. 2014.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s