The hunt for red tides: Deep learning algorithm forecasts shellfish toxicity at site scales in coastal Maine

Farmed and wild harvest shellfish industries are increasingly important components of coastal economies globally. Disruptions caused by harmful algal blooms (HABs), colloquially known as red tides, are likely to worsen with increasing aquaculture production, environmental pressures of coastal develo...

Full description

Saved in:
Bibliographic Details
Published in:Ecosphere (Washington, D.C) D.C), 2019-12, Vol.10 (12), p.n/a
Main Authors: Grasso, Isabella, Archer, Stephen D., Burnell, Craig, Tupper, Benjamin, Rauschenberg, Carlton, Kanwit, Kohl, Record, Nicholas R.
Format: Article
Language:English
Subjects:
Algal blooms
Algorithms
Application programming interface
Aquaculture
Artificial intelligence
Classification
Climate change
Coastal development
Decision making
Deep learning
Environmental changes
Eutrophication
Fisheries management
forecast
Forecasting
harmful algal bloom
Machine learning
Maine
neural network
Neural networks
paralytic shellfish toxin
Red tides
Shellfish
Toxicity
Toxins
Water quality
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Farmed and wild harvest shellfish industries are increasingly important components of coastal economies globally. Disruptions caused by harmful algal blooms (HABs), colloquially known as red tides, are likely to worsen with increasing aquaculture production, environmental pressures of coastal development, and climate change, necessitating improved HAB forecasts at finer spatial and temporal resolution. We leveraged a dataset of chemical analytical toxin measurements in coastal Maine to demonstrate a new machine learning approach for high‐resolution forecasting of paralytic shellfish toxin accumulation. The forecast used a deep learning neural network to provide weekly site‐specific forecasts of toxicity levels. The algorithm was trained on images constructed from a chemical fingerprint at each site composed of a series of toxic compound measurements. Under various forecasting configurations, the forecast had high accuracy, generally >95%, and successfully predicted the onset and end of nearly all closure‐level toxic events at the site scale at a one‐week forecast time. Tests of forecast range indicated a decline in accuracy at a three‐week forecast time. Results indicate that combining chemical analytical measurements with new machine learning tools is a promising way to provide reliable forecasts at the spatial and temporal scales useful for management and industry.
ISSN:2150-8925
2150-8925
DOI:10.1002/ecs2.2960