North Atlantic right whales are endangered, and the expanding blue economy—offshore energy, fisheries, and maritime transport—needs reliable information on where whales occur to avoid conflict. Autonomous underwater vehicles equipped with passive acoustic monitoring offer new data, but those records are statistically difficult to use: they contain many zeros, detection is imperfect, and the data have complex spatial and temporal dependencies.
The authors introduce DeepZIB, a deep zero-inflated Bernoulli model, which jointly models the latent presence of the species and the probability of detecting it while learning habitat relationships from heterogeneous covariates. They establish theoretical structural properties and use simulation experiments to show the model can recover underlying parameters and latent presence fields.
When applied to real passive acoustic monitoring data from the U.S. East Coast, DeepZIB demonstrates improved model adequacy and predictive performance. The model generates high-resolution, spatially and temporally varying presence maps, offering targeted, risk-aware guidance for managing blue economy activities in whale habitats.