Book Description

Marine renewable energy has the potential to provide clean, reliable power to coastal communities and offshore facilities. However, the effects that marine energy development might have on the environment are not yet well understood. One environmental risk of particular concern is that of collision between an animal and a marine energy converter, but conducting the requisite environmental monitoring to understand this risk has presented a challenge at marine energy sites around the world for several reasons. First, if collision does occur, it is likely to be a rare event, meaning that detection requires continuous monitoring over extended deployments. Second, there is no single sensor that can provide all of the necessary information, and a combination of active acoustic, passive acoustic, and optical sensors is required. Third, these sensors can rapidly accrue vast volumes of data (petabyte-scale), making it difficult to extract insight from collected data. Finally, waves and currents at marine energy sites complicate the deployment of any monitoring instrumentation. Integrated instrumentation platforms that combine sensors into a single platform can address some of these challenges, because they can provide all of the necessary data and reduce deployment complexity. However, operation of such a platform must meet three directives in order to be most effective: 1) avoid biasing animal behavior through the use of instrumentation, 2) reliably detect rare events, and 3) avoid collection of unmanageable volumes of data. In this thesis, it is demonstrated that it is possible to simultaneously meet all three of these directives. This is demonstrated using the Adaptable Monitoring Package (AMP), an integrated instrumentation platform that combines multibeam sonars, optical cameras, hydrophones, and an acoustic Doppler current profiler. Artificial illumination is necessary to collect data from optical cameras when ambient light is not available. However, this light can either attract or repel animals. To minimize these effects (e.g., meet directive 1), the AMP uses detection, tracking, and classification of targets in the multibeam sonar data to restrict the use of artificial illumination to periods when a target of interest is present and might be detectable by the optical cameras. Information about target presence is also used to limit data archival to periods when a target of interest is present and avoid curation of data that does not contain any useful information (e.g., meeting directives 2 and 3). To benchmark this capability, real-time target detection and tracking are used to limit data archival to periods when any target of potential interest is present during a deployment of the AMP in Sequim Bay, WA. The target detection and tracking approach was found to have a true negative rate of 0.99 (e.g., an estimated 1% of targets of interest were not recorded), but 45% of recorded data did not contain a biological target. To address this relatively high false positive rate, recorded data were used to train machine learning classification of tracked targets. Three machine learning algorithms, trained using varying parameters and features, were evaluated for this task. A random forest algorithm was found to perform best, and the resulting classification model was able to distinguish between biological targets (e.g., seals, fish) and non-biological targets (e.g., acoustic artifacts) with a true positive rate of 0.97 and a false negative rate of 0.13. This model was then implemented in real-time during a second deployment of the AMP and used to limit data acquisition to periods when biological targets were predicted to be present. The model achieved the same true positive rate and a false positive rate of 0.23 in real-time after re-training with site specific data. From these results, general recommendations are made for implementation of real-time classification of biological targets in multibeam sonar data at new marine energy sites. All active acoustic sensors used on the Adaptable Monitoring Package, including the multibeam sonar used for real-time classification, have operating frequencies above the upper limit of marine mammal hearing. However, high-frequency transducers can still produce sound at lower frequencies audible to marine mammals. A comprehensive evaluation of the acoustic emissions of four active acoustic transducers used on the Adaptable Monitoring Package was conducted to understand whether they might cause hearing damage or bias marine mammal hearing (e.g., violating directive 1). All four transducers were found to produce measurable sound below 160 kHz, the reported upper limit of marine mammal hearing. A spatial map of the acoustic emissions of each sonar was used to evaluate potential effects on marine mammal hearing if the transducer were continuously operated from a stationary platform. Based on the cumulative sound exposure level metric, the acoustic emissions from any of the the transducers are unlikely to cause hearing damage to marine mammals. However, the extent of audibility is estimated to be on the order of 100 m, and further research is needed to understand how this might affect marine mammal behavior. In sum, this thesis provides a framework for effective environmental monitoring that can be used to reduce the the uncertainty surrounding the environmental effects of marine renewable energy. Further, many aspects are widely applicable to the ocean instrumentation community. Automatic classification of fauna in multibeam sonar data had not been previously demonstrated, and has applications in biological research. The methods developed for evaluation of the acoustic emissions of active acoustic sensors allow for effective comparison between transducers, which can be used to inform sensor selection and government regulation of their use.