Computer Vision Tools for Marine Science Applications: Mapping and Measuring Marine Sedimentary Habitats and their Ecological Characteristics

Abstract

The onset of climate change and accelerating human activity has driven the need to closely monitor, assess, and measure impacts on marine ecosystems. Current tools and processes employed by marine ecologists suffer from a lack of the skilled labour force required to gather and process data with sufficient spatial and temporal coverage. Marine scientists need novel and cheaper tools to tackle these issues. Computer vision, the science of extracting information from digital images, enables marine scientists to do so with easily acquired image data. Reviewing the use of computer vision in marine science application over the past five decades, we showed that there is more to explore in using computer vision for improved processing and spatial coverage to satisfy marine science goals. We then focused our efforts on tackling three marine ecological problems to extend spatial assessment capabilities and provide new data for marine scientists. First, we looked at a multiscale computer vision approach to improve spatial coverage to estimate small biogenic sediment features across estuary sandflats. Using a quadcopter drone to capture image data at different altitudes, we demonstrate the potential of our multiscale depth and texture model to transfer information between spatial scales and estimate feature counts accurately. Secondly, we constructed a computer vision framework for rapid structural characterisation of soft sediment microtopography to capture the effects of changes in the macrofaunal communities that modify sediment topography through bioturbation. We formulated surface detrending models that operate on 3D depth data to extract millimetre scale changes in the sediment due to these animals. Lastly, we looked at 3D reconstruction to capture the distances between mussel structures on restored mussel beds to assess restoration efforts. We develop and assess a structure-from-motion based framework, with automated rescaling, for extracting metric measurements from the mussel bed reconstructions. This investigation highlights the first steps towards a completely automated mapping system to assist in the restoration of mussels in the Hauraki Gulf. This thesis demonstrates that the collaboration between the computer vision and marine science has many fruitful research possibilities with real-world impact. We hope the findings in our thesis will encourage deeper collaboration in the future.