New Technologies for Fisheries Assessments

Working alongside their NOAA collaborators aboard R/V Shimada, APL-UW investigators are testing artificial intelligence methods and a ship-to-cloud data processing pipeline against traditional methods to identify Pacific hake aggregations in acoustic data. The technologies they have developed to automate fishery stock assessments during routine surveys seek to increase the speed and consistency of data analysis and interpretation.

Principal Oceanographer Wu-Jung Lee and graduate student Caesar Tuguinay joined a 10-day leg of the Integrated West Coast Pelagics Survey when Shimada, a vessel capable of acoustic trawl surveys and oceanographic research, was conducting operations off the coast of Northern California.

Echosounders mounted on the vessel emit sound pulses into the water column below and measure the intensity and timing of returning echoes that bounce off objects including fish, plankton, and the seafloor. An echogram is a color-coded map of echo intensity. NOAA scientists aboard R/V Shimada are experts in the art of echogram analysis — labeling where along the survey line they see the distinctive acoustic signature of Pacific hake schools. Hake is the most abundant commercial fish stock on the Pacific coast, excluding Alaska, but the fish are difficult to distinguish in echograms because they can school at significant depth amid other fish species and they range over a large region with complex bathymetry. Though expert, scientists' manual interpretation of echogram data is time consuming, dependent on subjective labeling, and prone to human bias and unreproducible errors.

Efficiency, Consistency, Reproducibility, Interoperability

Machine learning algorithms (convolutional neural networks) are very good at finding complex features. In this case, Lee trained a neural network on datasets of echograms annotated by NOAA scientists that stretch back to 2005, teaching the model to recognize and classify the acoustic signatures of Pacific hake.

To maximize the efficiencies that AI can provide and achieve an echogram-to-biomass estimate in near-real time, the APL-UW team developed a modular, ship-to-cloud data processing pipeline. First, the echosounder raw files are consolidated and calibrated. Then the machine learning prediction computations are performed to identify regions of hake in the echograms. From here, data are transformed into a Nautical Area Scattering Coefficient (NASC), which is an acoustic proxy for fish number density. NASC is combined with net trawl catch data – length, sex, and weight of fish – to estimate abundance and biomass in the cloud. Data products are then available on cloud-hosted visualization dashboards that scientists onshore can access for web-based exploration and interpretation. All of this processing is completed on a suite of open-source software packages available to the community. 

To address the problem of disambiguating mixed species assemblages, the APL-UW team used opportunities between net trawls and CTD operations on R/V Shimada to deploy a profiling echosounder. The Kongsberg wideband autonomous transceiver was lowered from the ship deep into the water column. These new acoustic data, which are higher resolution than those from the ship's echosounder because of proximity, will be used to further train the models to recognize the differences among hake and other species in mixed assemblages.

Back at the Laboratory, Lee and Tuguinay are at work on methods to compare expert scientist annotations and model predictions at sea as the survey progresses. "Taking in their real-time annotations will coach the model progressively throughout the survey," says Lee. This could be a significant improvement in model performance because of observed year-to-year differences in the acoustic data.