In order to help restore elephant numbers, it’s vital that conservationists know how many there are in the wild. But counting elephants isn’t easy.
The playing field
The most common way – counting from a plan flying overhead – has its problems. Observers can miscount the elephants, whether due to tiredness or poor visibility. There are logistical challenges too: it’s costly, and there are difficulties with finding suitable runways and scheduling refuelling stops. When herds are in cross-border areas, it’s necessary to get permission, and this can be a complicated and time-consuming process.
Satellites, on the other hand, have several advantages. They can cover large areas in one go, meaning there is less risk of counting the same elephants twice. They don’t intrude on the animals’ habitat, nor are humans put at risk when collecting the data. They open up areas that might have been inaccessible before, including those cross-border areas that require extra bureaucracy.
The research project
With all this in mind, a team of researchers published a study in December 2020 that tested a way of counting elephants with the use of very-high resolution satellite images.
To analyse these images, they used another technology that has become an essential tool for ecologists: deep learning. Since 2012, a type of algorithm called a convolutional neural network (CNN) has revolutionised the field of computer vision. This has aided the detecting of wildlife in images that come from a variety of sources, including camera traps, aerial surveys, multibeam imaging sonar, and drones.
This isn’t the first time that very high-resolution satellite imagery has been used to count wildlife species. However, those studies would focus on landscapes that looked more or less the same everywhere. In these kinds of places, animals are easier to spot because they contrast strongly with their environment. In this study, the team wanted to know how feasible it was to count elephants living in an environment that’s more diverse and complex.
Their target was an elephant population in Addo Elephant National Park in South Africa, a place where the herds of elephants move between open savannah and diverse woodlands.
The results
The results of the project turned out to be very promising. The study demonstrates, for the first time, that it is possible to automate detection and counting of African elephants in heterogenous landscapes using a combination of state-of-the-art satellite remote sensing and deep learning technologies. The method shows lots of promise as an effective new wildlife surveying technique.
Through the creation of a customized training dataset and the application of a CNN, the accurate detection of elephants in satellite imagery has been automated to a level that matches human detection capabilities. Even elephant calves were accurately detected, despite their absence in the training dataset.
The success of the model to detect elephants outside of the training data site (and also using data from lower resolution satellites) demonstrates its generalizability. This is a very promising feature, as even small amounts of training data from other localities or satellites would further increase accuracy.
The study also demonstrates that speeding up the identification of species by automating detection can allow for large-scale application of satellite-based wildlife surveying. A detection process that would formerly have taken weeks can now be completed in a matter of hours.
Finally, observer variability means that errors in human-labelled datasets are inconsistently biased. In contrast, false negatives and false positives in deep learning algorithms are consistent and can be rectified by systematically improving models.
To find out more about this study, you can read the original journal article here
