Antarctica: ENEA counts Antarctic penguins with supercomputer

ENEA will use for the first time artificial intelligence in Antarctica to count and classify Adelia penguins living near the Italian Station Mario Zucchelli applying a method based on drones, artificial intelligence (AI) and the ENEA supercomputer CRESCO.

“The accuracy of the population count was 97 percent for adult penguins and 89 percent for chicks, which is a high level of accuracy,” explained ENEA researcher Riccardo Scipinotti, current expedition leader at the Italian-French Concordia base on the Antarctic Plateau, more than 3,000 meters above sea level and 1,200 kilometers from the coast.

These results were obtained by analyzing orthophotos acquired at the two penguin coves closest to the Zucchelli base (Edmonson Point and Adelie Cove), which count about 4,000 and 10,000 individuals respectively. To enable researchers to estimate the colony's reproductive efficiency -one of the indicators of animal health and climate change- the count is repeated at two precise times of the austral summer, mid-November, the hatching period, and mid-January, when the chicks are free to move around the penguin cove.

“Specifically, the count was obtained employing special drones technically suitable for flying in low Antarctic temperatures and over vast remote areas. In order to film some colonies without disturbing the animals, it was necessary to exceed an altitude of 50m,” said Scipinotti.

The images are first processed by the CRESCO supercomputing cluster, active since 2017 at the Zucchelli station, which returns a high-resolution (about 1cm/px) georeferenced orthophoto of the pinguinaia directly in the field.

In fact, the excellent resolution of the image enables to identify both the adult penguins, which are about 60cm in size, and the chicks, which at about 2 months of age barely reach a height of 30cm.

The orthophoto is then processed by a neural network specially trained to recognize the penguins in the image, distinguishing them between adults and chicks.

“The ENEA CRESCO supercomputer at the ENEA Portici Center was used to train the neural network, fed with about 400 drone images of more than 3,000 penguins, taken under different exposure conditions and positions during the last eight research expeditions,” explained Samuele Pierattini, of ENEA Division for the Development of Systems for Computing and ICT.

“Compared to manual spot counting by field personnel, this method is faster and allows monitoring of remote areas difficult for humans to access, while reducing disturbance to animal species,” Scipinotti pointed out.

The next step will be refining the network training with data acquired during the expeditions so as to increase its reliability and to specialize it also to detect and distinguish Emperor penguins, another species that colonizes near the Mario Zucchelli Station.

In addition, new AI technologies will be employed to build a computationally 'light' neural network that can be installed inside the drone's electronics and provide real time count of penguins.

“This would make it no longer necessary to export raw data from Antarctica to be processed in laboratories, as already complete information would be available to assess and study the numerical status of the penguin population,” Pierattini concluded.

The 40th expedition of the National Program of Research in Antarctica (PNRA), funded by the Ministry of Universities and Research (MUR) and managed by the National Research Council (CNR) for scientific coordination, ENEA for logistic planning and organization of activities at Antarctic bases, and the National Institute of Oceanography and Experimental Geophysics (OGS) for technical and scientific management of the icebreaker ship Laura Bassi, is currently underway.