v.5.0 is out Now!
Reading time:

Neural network to calculate the population of Saiga antelope in collaboration with WWF

SPH Engineering
October 3, 2019

WWF has joined forces with SPH Engineering to develop an automated system of calculating populations of Saiga antelope Saiga tatarica using data obtained with UAVs.

In 2018-2019 WWF researched methods of counting Saiga antelopes with the help of UAVs. In June 2019, an aerial photo shoot of the animals was performed in Chornye Zemli, Mekletinskii, and Stepnoi nature reserves.

An analysis of the pictures taken during the photoshoot performed by wildlife experts revealed several seasonal aggregations and their composition by sex and age. This kind of manual analysis of aerial photography is a labor-intensive process that requires the involvement of highly-qualified professionals. In order to optimize the process of analyzing such images, WWF called in the specialists from SPH Engineering - one of the leaders in the field of UAV software development and digital analysis of photographs. They developed and trained neural network algorithms that could recognize all antelopes on a given image and identify their age and sex.

An example of aerial photography of the antelopes scattered across a diverse landscape

At the moment, experts are comparing the results of the manual and automated analyses of the images to determine whether any changes in the neural network’s detection strategy or learning conditions are needed.

“It is always a pleasure to see modern technologies assisting in the preservation of wildlife, rather than causing it harm,” says Alexei Yankelevich, the R&D director at SPH Engineering. “We achieved the goal - our neural network algorithm allowed the specialists from WWF to do their work much more efficiently. The task turned out to be more challenging than anticipated. The photoshoot was carried out at a high altitude in order not to scare the animals with the sound of drone engines. Therefore, there were very few dots per animal on a picture. This compelled us to be extremely careful with the creation of the learning dataset and the processing pipeline, so the quality of detection and classification of the animals would be appropriate.”