Resource Efficient Data Transmission through Pattern Recognition

In the research field of animal behaviour, electronic devices which track an animal's behaviour with various sensors, so-called bio-loggers, have become omnipresent. Therefore, it is unsurprising that optimisation questions like data quality, size, weight, memory, transmission rate and energy efficiency have become topics of high relevance. Development in all those fields holds promising progress and enables ever-new solutions for one another. In this thesis, I take on the task of decreasing the energy consumption of bio-loggers. I use pattern recognition from machine learning, in the form of decision trees, to only send information once a specific pattern has been recognised. I describe how this can help to significantly decrease the amount of information transmitted. I developed, deployed, and tested the decision tree models on the WildFi tag, a state-of-the-art bio-logger which excels in the aforementioned qualities. In particular, this thesis revolves around which possibilities to save power arise from pattern recognition, whether a decision tree method suffices for this task, and what impact omitting sensors can have on recognition accuracy and power consumption. I discovered that the potential for saving energy does not depend only on the frequency of a target behaviour. Indeed, there is evidence that different behaviours have varying sets of sensors required to describe them with maximised accuracy. Thus, not every behaviour of an animal necessarily requires every sensor to be active for detection. My findings constitute an interesting potential for further exploration to use pattern recognition on bio-loggers for informed and controlled transmission of information.