2023 Author: Bryan Walter | [email protected]. Last modified: 2023-05-21 22:24
Researchers from the UK and the Netherlands have taught the neural network to recognize and count elephants in satellite images. The algorithm works not only with images on an empty surface, but also, for example, if there are many trees on it, and the accuracy of its work is approximately equal to that of a human, the authors of the preprint in bioRxiv say.
Zoologists use aerial and satellite imagery to monitor animal populations and their migration. This makes it possible to conduct observations remotely, as well as cover vast territories, spending little resources. In most cases, researchers have to manually label the animals in the images, so it takes a long time. This stage can be automated, but it is technologically difficult, because often in the images, in addition to animals, there are other objects, for example, shrubs, trees or stones.
In recent years, this problem has often been solved with the help of neural networks, which cope with the task better than analytical algorithms that distinguish objects by threshold values of size and color. Researchers led by Isla Duporge of the University of Oxford trained a neural network to recognize elephants using high-resolution satellite imagery.
They used data from WorldView-3 and -4 satellites with a resolution of 31 centimeters per pixel - the highest among commercial satellites providing data. The photographs captured the territory of the Eddo Elephant National Park in South Africa. This area has many low and tall plants and is home to about 600 elephants. They often cover themselves with mud to cool, making them uneven and difficult to detect.
The authors used archived images for the period from 2014 to 2019, processed them with a panchromatic fusion algorithm to obtain greater clarity, and cut them into 600 x 600 pixel fragments that are compatible in size with the neural network. They manually annotated the pictures and identified 1,125 elephants on them. As an algorithm for detecting animals, the researchers chose the Inception ResNet convolutional neural network, pre-trained on a dataset of various everyday objects COCO (pre-training allows you to quickly retrain the neural network for a specific task).
The territory of the park, as well as examples of homogeneous and heterogeneous landscapes (right)
The researchers tested the neural network, comparing its results with the results of 51 volunteers who manually tagged the data. They chose the F2 parameter for the assessment, which takes into account frequently used metrics, such as completeness (how many real elephants the algorithm found) and accuracy (how many objects allocated by the algorithm are elephants) of the results, but focuses on false negative results, because it is easier for the checking person weed out incorrectly recognized elephants in the pictures, rather than look for unrecognized ones. Also, the authors separately compared the results on the images with a homogeneous (steppe) and heterogeneous (many trees or bushes) landscape.
The neural network received a result of 0, 778 for heterogeneous areas and 0, 73 for homogeneous areas, and for volunteers, the result was 0, 8 and 0, 776, respectively. The authors note that the main obstacle to improving the quality of recognition remains the price of satellite images, which in their case was $ 17.5 per square kilometer for archived images and $ 27.5 for new ones.
Recently, British scientists have successfully used satellite imagery to find new emperor penguin colonies, finding 11 new colonies and increasing the number of currently known nesting sites to 65.
