Africa is home to a vast array of species. However, climate change and other factors threaten their existance, so it is important to scientifically monitor these organisms. Previous solutions to the monitoring task have included counting species from aircraft, which has several limitations. Satellite imagery, on the other hand overcomes many of the problems that a manual and phsycial approach takes. Deep learning has been applied to counting whales, and elephants from space, but the difficulty increases as the animal becomes smaller as current satellites don’t have a fine enough focus. In this paper the authors demonstrate a deep learning approach that locates and counts large groups of animals like the wildebeest.
Key Takeaways
The f1 score of this model is 84.75 percent, so the model performs quite well on the task of identify wildebeest in an environment that creates confusion with a variety of wildebeest-esc objects. From above, bushes and termite mounds can look similar to wildebeest, which adds confounding objects to the task of identifying wildebeest in this heterogenous landscape. The model transfers well to years it hadn’t seen before, trained on all the data but 2020 and testing on 2020 the model has the same performance. The same performance for 2015.
Methodologies
The satellite imagery that authors use looks at the Serengeti National Park and the Masai Mara National Reserve on the border of Tanzania and Kenya. The model they used is call U-Net, which is a specific architecture of a convolutional neural network (CNN). The UNet framework has compresses and decompresses the image as it applies the traditional CNN process. As U-Net decompresses an image it combines the same sized image from the compression steps. U-Net was originally built for segmentmentation of biological images, so in this situation it assigns a probability of being a wildebeest to each pixel. Therefore, the authors use kmeans to find the possible centers of wildebeest.
Citation
Wu, Z., Zhang, C., Gu, X., Duporge, I., Hughey, L. F., Stabach, J. A., … & Wang, T. (2023). Deep learning enables satellite-based monitoring of large populations of terrestrial mammals across heterogeneous landscape. Nature communications, 14(1), 3072