Combining real-time and historic marine mammal information for sonar risk assessment forecasting and backtesting

Incorporating (near) real-time detection data into active sonar risk assessments has been difficult due to challenges in collecting the data and fusing multiple sensing modalities. A mixture model is developed in this work which splits marine mammals in the area into known and unknown mammals according to the detections made by sensors. The spatial-temporal coverage of the same sensors is also incorporated since they provide additional information on where the mammals are not located when no detections have been made. The detection data is used to conditionally update the probability of harm to the known and unknown marine mammals due to active sonar and the probability of where to find these mammals. The total number of harmed mammals are aggregated, and worst-case scenarios are quantified using risk measures from mathematical finance. These risk measures also provide a way to backtest the underlying models the risk framework depends on. An example backtest against the underlying mammal density model is shown using the Value-at-risk measure with simulated detection data. This work is a promising further step towards leveraging real-time monitoring data for minimizing the risk of harm to marine mammals.