Abstract
Coral reef ecosystems are rapidly changing due to global climate change, with marine heat waves and human stressors exerting significant pressure. This has led to the decline of marine biodiversity, most famously in corals, a keystone species that supports global ocean ecosystems. While there have been advancements in assessing benthic substrate biodiversity in reefs, the evaluation of motile marine wildlife biodiversity has lagged, with potential biases from in-situ disturbances. NASA's airborne fluid lensing technology, deployed on autonomous aircraft, has enabled 3D imaging of coral reef ecosystems at the cm-scale to depths of 20m, aiding in benthic biodiversity assessment without wave distortions. However, this technology removes transient artifacts, including motile marine life, which is crucial for quantifying biodiversity in shallow marine ecosystems. We develop a new airborne fluid lensing-based marine wildlife detection algorithm to characterize moving underwater wildlife using raw fluid lensing data and solutions. The detector assesses marine animals in time series imagery, deriving animal planform area, velocity, and abundance across depths of ∼0-10m, at cm-scale resolution, in complex coral reef ecosystems over dozens of square kilometers. The algorithm's accuracy and performance were evaluated using a consensus-based technique in a single-blind experiment. The datasets span three ecologically distinct coral reefs in Guam. Results indicate the algorithm detects marine wildlife three times more efficiently than humans, with 6% less precision. Its inconsistent recall, dropping to 10% at its worst, indicates sensitivity to environmental conditions. These high quality results could greatly lower the cost of dataset curation for future machine learning methods.