In collaboration with researchers at the University of St. Andrews, Zoological Society of London, University of Utah, Northern Illinois University, University of Cape Town, University of Roehampton and Bawangling National Nature Reserve, we have developed machine learning algorithms to detect the vocalisations of the world's rarest primate, the Hainan gibbon. Futhermore, together with researchers at the City University of New York, New York Consortium in Evolutionary Primatology, University of Antananarivo, Institut International de Science Sociale, and Centre ValBio, we have extended our algorithms for the automated detection of critically endangered black-and-white ruffed lemurs in Madagascar.

Photo credit: Jessica Bryant, Zoological Society of London

In this research, we aim to explore the capabilities of compressed sensing (CS) techniques for data compression to achieve the development of real-time bioacoustic monitoring. CS can be used to compress recorded data on recording devices, and then the compressed data can be efficiently transmitted over a wireless network or via a slow internet connection in remote areas. For this to be useful, there should be a means to reconstruct the data on a remote server to obtain animal vocalization predictions using deep-learning techniques.

Photo credit: Jessica Bryant (Hainan gibbon), Francesco Veronesi (Thyolo alethe).

Together with Professor Tiago Marques (University of St. Andrews and University of Lisbon), Dr Kalliopi Gkikopoulou (University of St. Andrews) and Tomas Gueifao, we are developing machine learning algorithms that can detect vocalisation events of beaked whales. We are also developing algorithms to detect the vocalisations of narwhals, a project in collaboration with Professor Tiago Marques, Dr. Carl Donovan (University of St. Andrews) and Carolina Marques. These algorithms help us better understand marine mammal vocalisations. In collaboration with Dr  Guilherme Frainer and Sea Search, we are developing machine learning algorithms for the automated detection of engandered humpback dolphins in South African waters.

Photo credit: Rene Swift, DTAG (Johnson & Tyack, 2003).

In the field of deep learning, it is well known that a large amount of data is required when developing algorithms for specific tasks such as classifying images. Just like how we learn better when we have more examples and experiences, deep learning models become more accurate and reliable with a greater variety and quantity of data. More data allows the model to generalise better, making better predictions and decisions when faced with new, unseen situations. Data acquisition in ecological applications presents unique challenges due to factors such as remote and inaccessible environments and endangered species to name a few.  A possible solution to overcome the limitations of having limited or insufficient data is to generate synthetic data that captures the underlying patterns and characteristics of the real data. The synthetic data introduces diversity and equips the model to become better at generalising and handling variations and complexities in real-world scenarios. To this end, the objective of this project is to generate synthetic audio data (in the form of image representations) of the Hainan gibbon call for improved bioacoustic classification. 

Neural networks are a powerful tool used for predictive modelling in multiple different environments. In the field of ecology, it can be used to predict and/or analyze animal behaviours, automating tasks that usually take hundreds of man-hours to complete. However, these neural network models are often very large, requiring a lot of storage space and energy to compute its predictions. Thus, we are looking at ways to compress these models as much as possible to minimize storage and energy requirements while maintaining the accuracy of predictions as far as possible.

The African Penguin is an endangered species of penguin found only off the coast of Southern Africa. The aim of this project is to build a deep learning algorithm that can accurately predict penguin posture and depth from monocular video feed. This will allow us to classify penguin behaviours and better understand interactions between penguins. This will allow ecologists to gain better insights into penguin colonies and aid in conservation of this rapidly declining species.

Chinstrap penguins, which breed on the Antarctic Peninsula, possess remarkable diving capabilities. They undertake deep dives during which they engage in foraging activities, primarily feeding on krill. The objective of this project is to deploy bio-loggers, specifically accelerometers, and utilize deep learning techniques to automatically detect and quantify the number of prey capture events performed by the penguins.

Project led by Dr. Chris Oosthuizen, realized by Dr. Stefan Schoombie (University of Cape Town, South Africa).

The objective of this project is to create an algorithm that automatically identifies hawksbill sea turtle behaviors using data from bio-loggers. To achieve this, a transfer learning approach will be employed, leveraging a pre-existing model developed for green turtles (the Vnet, see image). While transfer learning methods are commonly used in image classification to train models on limited datasets, their application in the field of bio-logging remains unexplored. This project aims to bridge this gap and assess the effectiveness of transfer learning methods for behavioral identification in sea turtles.

Project in collaboration with Dr. Damien Chevallier (CNRS BOREA, France)

Just as a human would use contextual information to identify species calls from acoustic recordings, one unexplored way to improve deep learning in bioacoustics is to provide the algorithm with contextual metadata for each recording, such as time and location. In this project, we tested different methods to incorporate contextual information to deep learning acoustic classifier. As a first case study, a multi-branch convolutional neural network (CNN) was developed to classify 22 different bird songs using spectrograms as a first input, and spatial metadata as a secondary input. A comparison was made to a baseline model with only spectrogram input. A geographical prior neural network was trained, separately, to estimate the probability of a species occurring at a given location. The output of this network was combined with the baseline CNN. As a second case study, temporal and spectrogram data was used as input to a multi-branch CNN for the detection of Hainan gibbon calls, the world's rarest primate.