Human advancement may be threatening nature, but it also offers a way to preserve it.
It is no secret that human activity is harming the natural world. Poaching, habitat loss and pollution are only a few examples of the ways in which species and ecosystems are under threat. Yet protecting nature isn’t without hope, and conservation plays a vital role in offsetting these actions. However, resources are limited, so to ensure that they are not misallocated it is essential to have reliable information on where species are and how many are in a region. This is significantly easier said than done. Traditional methods of animal tracking involve establishing camera trap grids, searching for tracks or using aircraft for aerial surveys. These methods risk being invasive, logistically complex and don’t guarantee a full picture of animal distribution. They can also be error-prone, particularly over large areas.
For the first time, satellite images and machine learning have been used together to recognise and count animals in complex landscapes. By studying African Savannah elephants in a diverse environment, a team led by Dr Isla Duporge from the University of Oxford has explored the potential of this technology as a new method to aid wildlife conservation.
“Traditional methods of animal tracking can be error-prone, especially over large areas”
Durporge and her team used high-resolution satellite images to train a Convolution Neural Network (CNN) model. These models have multiple uses outside of conservation, including facial recognition. The training images consisted of aerial views of elephants in South Africa’s third-largest park, Addo Elephant National Park, where over 600 elephants spend their days roaming between open grassland and densely packed shrubs. African elephants are an excellent test subject for such a study as their size makes them considerably more straightforward to find amongst diverse landscapes than smaller species. This does not mean that training the model was simple. Elephants will take on different forms during the day through activities such as sleeping, foraging, and walking. During the hot hours of the day, they will commonly use mud baths to cool down, so the model also had to be adapted to be able to recognise elephants whether grey or covered in coppery-red clay.
The resulting CNN model was able to recognise and count the elephants with a similar degree of accuracy to human observers. Most excitedly, this model could be generalised to other elephant populations as displayed when tested outside of its training area using lower-resolution satellite images from the Maasi Mara in Kenya. Here, it was able to recognise over half of the elephants against the new backdrop. The team predicts that with a small amount of additional training, the model would be just as powerful in recognising elephants in this new environment.
“The model was able to recognise and count the elephants with a similar degree of accuracy to human observers.”
Using satellite images avoids the issues previously stated which arise when using traditional methods . The WorldView 3 and 4 satellites used in this study are able to collect high-resolution imagery from over 5,000km2 in a matter of minutes, allowing for vast amounts of information to be collected without disrupting the animals. Another benefit is that monitoring can occur across borders and in areas with restricted accessibility. Counting animals in such a large collection of images is a laborious task. Machine learning can be used to automate it, processing information far quicker than a team of human observers. This frees up time for researchers to use the information to develop conservation strategies that would otherwise be spent collecting raw data.
Previous studies have already used satellites imagery and machine learning to recognise animals from space. However, these studies focus on animals that stand out against simple backdrops of the blue sea or white ice-sheets. Unfortunately, many animals live in more complex habitats, residing among foliage or travelling between a range of environments. This study is therefore an impressive new development in conservation science and is just the beginning of a new method of working. Satellite imagery resolution is continuing to improve, allowing for smaller animals to be observed in higher levels of detail. Later this year, a new constellation of six satellites is going to be launched that will provide larger geographical coverage, allowing for a broader range of species to be studied in higher resolution. While not every species may be able to be counted from space due to their size or habitat, this use of satellite imagery and machine learning can certainly be used to improve conservation efforts for a number of threatened species.
Written by Sophie Teall and edited by Shona Richardson
Sophie’s thoughts…The results of this paper are certainly exciting. Traditional methods of large scale data collection are no small feat, even before considering all the paperwork and bureaucracy that can be involved when studying animals that move between protected areas and countries. That having been said, satellite imaging and machine learning can՛t be used to track every animal of interest, as many are simply far too small or live under dense tree canopies. There is already a bias in conservation and research to focus on species viewed as charismatic, which often involves large mammals. As wonderful as these animals are (elephants are definitely one of my favourites), these popular animals make up only a tiny fraction of the species on the planet. Many threatened animals that are vital to ecosystems are significantly smaller, so I hope that as larger species become easier to track the literature divide between these groups won’t grow.
Sophie is a final year undergraduate biology student, specialising in Zoology. Find her on Twitter @sophteall and LinkedIn @Sophie Teall.