Automated acoustic monitoring of ecosystems

Abstract

Natural habitats are being impacted by human pressures at an alarming rate. Monitoring these ecosystem-level changes often requires labour-intensive surveys that are unable to detect rapid or unanticipated environmental changes. In this thesis, we explore how recording and analysing the sounds of an environment provides a tractable solution to scalable, fully automated ecological monitoring. First, we tackle the problem of autonomous data collection, and develop a device which is able to continuously collect and remotely transmit data from field sites over long time-periods. We then move to the automated analysis of eco-acoustic data, and exploit a learned acoustic feature embedding to achieve accurate monitoring of ecosystem health across multiple spatial and temporal scales. We demonstrate that an unsupervised approach using the same acoustic feature space allows automatic identification of anomalous sounds, including hallmarks of illegal activity such as gunshots and chainsaws. Functional real-time ecological monitoring requires significant computational infrastructure, and we detail the open-source design and implementation of SAFE Acoustics, an eco-acoustic monitoring network in the tropical rainforests of Borneo. Within the ecosystems we study, species movement and behaviour are constrained by habitat structure and connectivity. However, investigating how topology influences a complex network’s behaviour is not a challenge unique to ecology. We investigate the link between structure and function in a model system, the mouse brain, and find that inter-regional axonal connectivity is closely related to the intrinsic dynamics of individual brain areas. Similar studies inspecting the dynamics of data from large-scale ecological monitoring networks may provide a fruitful avenue for further explorations.