Automated detection of wildlife in proximity to marine renewable energy infrastructure using machine learning of underwater imagery

Abstract

The use of automated image analysis for monitoring wildlife interactions with marine renewable energy infrastructure can vastly reduce the time required to extract usable data from underwater imagery compared to manual expert processing. We present a novel industryready image processing workflow for automated wildlife detection developed using 1000+ hours of underwater video footage obtained by Nova Innovation Ltd. from their operational tidal stream turbine array at Bluemull Sound in Shetland, Scotland. The workflow includes object detection through advanced image analysis, image classification using machine learning, statistical analyses, and automated production of a summary report. Blind tests were undertaken on a subset of videos to quantify and iteratively improve the accuracy of the results. The final iteration of the workflow delivered an accuracy of 80% for the identification of marine mammals, diving birds and fish when a three-category (wildlife, algae, and background) classification system was used. The accuracy rose to 94.1 % when a two-category system was used, and objects were classified simply as ‘target’ or ‘non-target’. The accuracy and speed of the workflow can be improved through expanding the initial training dataset of images with different species and water conditions. Application of this workflow significantly reduces manual processing and interpretation time, which can be a significant burden on project developers. Automated processing provides a subset for more focused manual scrutiny and analysis, while reducing the overall size of dataset requiring storage. Auto-reporting can be used to provide outputs for marine regulators to meet monitoring reporting conditions of project licences.