2023_programme: Time and cost gains enabled by Machine Learning for Environmental Impact assessment.

Back


  • Session: 15. Towards Automatic Target Recognition. Detection, Classification and Modelling
    Organiser(s): Johannes Groen, Yan Pailhas, Roy Edgar Hansen, Jessica Topple and Narada Warakagoda
  • Lecture: Time and cost gains enabled by Machine Learning for Environmental Impact assessment.
    Paper ID: 1926
    Author(s): Semião Mariana, Hoofd Bénédicte, Cruz Erica, Almeida Diana, Vieira Susana, Vaz Guilherme
    Presenter: Cruz Erica
    Abstract: An Environmental Impact Assessment (EIA) is a process aiming to assess a priori the impact that a large scale project would have on an ecosystem, to propose mitigation measures to minimise potential impact, and also to monitor the effectiveness of the measures during the execution of the project. The process often lasts up to 4 years. It is a mandatory step for the deployment of any new projects at sea such as the construction of a harbor terminal, the installation of an offshore wind farm or an aquaculture cage. One common threat that such projects pose is an increase of underwater noise. This topic is particularly important when assessing the impact on marine mammals and fish because they use underwater sound to interact for feeding, mating, or socializing. \nOne of the methods to assess the impact of underwater noise on marine mammals is to estimate the presence/absence of animals in the specific region of the project using passive acoustic monitoring. This process is usually expensive: it requires the expensive installation of hydrophones, the manual recovery of the recorded data and the equipment with divers and ships, and finally the long manual analysis conducted by an expert to distinguish the different types of dolphins’ vocalizations.\nThis paper evaluates the time and cost gains that a Machine Learning algorithm can bring for the detection of acoustic sources of a specific site. A Deep Learning ensemble model is trained to detect dolphins in a coastal environment in Portugal, using a manually labelled dataset. The paper establishes the minimum requirements in terms of training dataset to allow for an automated, accurate, and fast analysis of the dolphins’ behavior in the area. The requirements consider the size of the dataset, but also the class balance and data processing required for the analysis. It was found that for the analysis of the used dataset, the labelling efforts can be reduced by a factor of 15.
      Download the full paper
  • Corresponding author: Ms Bénédicte Hoofd
    Affiliation: blueOASIS
    Country: Portugal
    e-mail:

Back