Automated conservation with deep learning AI for camera-trap identification of species and individuals (Ref: 4659)

点击翻译按钮获取中文摘要

The project will use deep learning approaches in conservation ecology for the automated identification of species or individuals from camera trap images. The topic promising to scale up conservation initiatives and to deal with a crucial bottleneck of conservation research: identification of wild animals which is resource- and time-consuming. It will also broaden the field of applied mathematics and computer science to further research into the discipline's power to solve modern world issues such as the well-documented biodiversity loss crisis. The project will begin with a focus on chimpanzee identification at three levels. First, using chimpanzee camera trap images, we will develop software to recognise the species within images using deep learning AI. The coding platform we will use is PyTorch, which has been previously used for such AI training for other species (Ramirez, 2022; Kholiavchenko, 2022; Lamba & Cassey, 2019; Wearn, Freeman & Jacoby, 2019). This work will draw on and develop machine learning approaches, specifically deep neural networks. Second, software to identify individuals at a species-level will be created, for example using facial recognition and accounting for the chimpanzees' ability to move their ears. Each chimpanzee will receive a unique identification number. Third, the software will be trained to recognise symptoms of visual diseases, such as leprosy, which has since not been attempted. Further extensions are also possible, such as print-recognition to identify leopard individuals in camera trap images. Existing contacts to a reserve in South Africa by the supervisors could lead to interesting appplications in their leopard conservation programme. Automating the identification of leopards and individuals at a species-level in readily available camera-trap images would propel their conservation within the park. The development of this software will also be on PyTorch, using 'flat' patches of print to distinguish between individuals.After the algorithms have been adequately trained (for example to match human volunteer accuracy at 99.6%, Sreedevi, 2022), wider applicability includin Population Projection Matrix (PPM) modelling and Value of Information theory (in collaboration with the University of Queensland) can then be explored to reveal more information on: - The optimal number of camera-traps (and time used) to adequately capture the population at a minimum cost/camera deployment rate, i.e. at what time are camera-traps simply recording the same individuals?- The distributions of species across reserves (including stage-structured distributions).- Tracking individuals of a species without the use of invasive techniques such as GPS collars by their recorded movements past the camera traps.This pioneering research has extremely strong links to real-world applications that extend further than the scope of this project. The addition of deep learning AI to identifications through camera-trap datasets not only promises to decrease the time of human-effort on such a task, it facilitates the draining of a vast bottleneck in conservation research. Thus, harnessing the power of computer science to aid in the conservation of threatened species.