Artificial intelligence serving the nature

Science and Research

Air pollution, extreme drought, loss of biodiversity or melting glaciers. The climate crisis has significant impacts on the planet and its life. That is why today we increasingly turn our attention to modern technologies as a potential solution – and artificial intelligence (AI) is one of them.

Source of photo: Kostas Papafitsoros 

Authors: Lukáš Adam and Lukáš Picek

AI is a useful tool especially when it comes to data analysis. AI-based algorithms can process large amounts of data and spot patterns or anomalies thanks to neural networks that have a more powerful processor than the human brain. This capacity is most often used for language models (used e.g. in the popular ChatGPT) and image processing. The latter in the context of ecology is our topic.

Image data originates from a number of sources. What are some examples? Satellite images are used to analyze changes in forest cover. Data on the loss of vegetation or changes in its color help conservationists detect loss of biodiversity. And our personal favorite – even wildlife can be monitored through the analysis of photos taken by drones, camera traps or ordinary cameras.

AI-powered wildlife monitoring

One of the uses of AI for wildlife monitoring is so-called re-identification, a term used for recognition of individuals. This field has gained importance in recent years due to the critical need for effective wildlife conservation and management strategies. Re-identification can provide important data on population size and population dynamics, such as birth and death rates, which are critical to understanding population health and developing conservation strategies. By re-identifying individual animals, scientists can also study movement patterns or help develop and evaluate conservation interventions.

Wildlife re-identification has traditionally relied on tracking technologies placed directly on the animals. For example, GPS collars can monitor the movement patterns of a group of animals over time and reveal their home ranges, migration routes and habitat preferences. Although these methods are effective, they are laborious, time-consuming, and handling of animals can be stressful for them.

Another approach is the use of genetic techniques which play a vital role in the re-identification of wildlife. Researchers can take samples of their feces or fur to analyze the DNA. This allows accurate re-identification of individuals without physically capturing or handling them. The key disadvantage? High price.

Gentle camera traps instead of collars or DNA samples

A suitable approach to re-identification is the use of non-invasive methods such as camera traps. These strategically placed cameras are activated by movement and capture images of animals in their natural habitat. These images can later be analyzed to identify individual animals based on distinguishing unique marks such as coat patterns, scars or other physical characteristics. Traditionally, these photos are analyzed manually. In recent years, AI has emerged as a powerful tool in this process by enabling more efficient and accurate identification of individual animals from large datasets.

There are several research groups in the Czech Republic that are interested in the application of AI to wildlife monitoring. Let’s look at two of them – one is located in our AI Center at CTU, the other at the Faculty of Applied Sciences of the University of West Bohemia in Pilsen (FAV ZČU).

Photographing sea turtles underwater. Source: Galini Samlidou

Chasing the sea turtles on Zakynthos

The group in our center is led by Lukáš Adam. His story with wildlife conservation began in Berlin. When his research stay was coming to an end, his Greek colleague Kostas Papafitsoros mentioned going to his homeland every year to help as a volunteer for the sea turtle monitoring at the Archelon, the Sea Turtle Protection Society of Greece. When he revealed that snorkeling is part of the job, Lukáš did not hesitate for a second and went with his colleague to Zakynthos where he spent an amazing two weeks on the popular island participating in conservation activities including raising public awareness and monitoring sea turtle nests. 

During 13 years of diving, his colleague Kostas collected over 50,000 photos of 612 individual turtles. Every time he encountered a turtle, he needed to identify it, which was very time-consuming. That is why he approached Lukáš last year with a proposal to develop an automatic system that would help him recognize individual turtles. Both of them started working on the project which soon brought the first results.

How to identify a turtle step by step

Re-identification begins with the detection of the turtle's head. For this, the researchers used a model originally trained to detect vehicles, road signs or pedestrians and retrained it to recognize turtles. They focused on the head, where the characteristic plates are located. Its shapes do not change throughout the turtle's life and can therefore be used to identify a particular turtle. By cropping the head, they removed the noisy background that can confuse re-identification software.

They then used additional AI models to automatically extract features from each image and compare them to other photos in the database. When these features were similar in two different images, the images were likely to contain the same individual.

Turtle identification based on a comparison of ten automatically generated features.

They published the edited set of 8,000 images of 400 individual turtles under the name SeaTurtleID. Anyone can use it and test their ability to identify individual turtles. Publicly available databases like this allow scientists to train algorithms to re-identify wildlife. You can also find their paper on arXiv to learn more about it.

And what is their big dream? Lukáš and Kostas want to create an interactive website focused on citizen science, where people could send their photos of turtles and get more detailed information about them, such as how many times it has been taken and where. We wish them good luck!

Biologist and the beast 

The second group we wish to introduce is led by Lukáš Picek. Together with his international team at the  University of West Bohemia, Lukáš has already developed several useful applications in ecology. In collaboration with the Danish Mycological Society, he created and launched a mobile and web app for the recognition of hundreds of mushroom species. He also contributed to a sophisticated system identifying thousands of snake species to help prevent snakebite mortality. He also built a system to identify wildlife species from camera traps in the Serengeti National Park in Tanzania, achieving first place out of 811 competitors.

Lukáš is the lead researcher at the CarnivoreID project. Its goal is to research and develop new AI technologies for the analysis of animal photos from camera traps. The future technology will have the potential to significantly increase the efficiency of the annotation process, shorten the reaction time to urgent situations (for example, the occurrence of a conflict species in new areas or an injured animal), improve and simplify the identification of individuals, and make more accurate estimates of endangered species populations.

3D model of a lynx from a camera trap. Source: Lukáš Picek

This article was written for Nový prostor, a magazine helping people in distress. 

Responsible person Ing. Mgr. Radovan Suk