Predicting estuarine algal blooms using artificial neural networks

Algae proliferate when favourable biological, chemical and physical conditions are present. This study uses data from the Berowra Estuary a tributary of the Hawkesbury River, NSW. Algal blooms in this estuary are a regular feature of seasonal cycles and develop in response to non-periodic disturbances. Toxic algae, mostly Dinophyta, have led to fish kills and estuary closure. Most non-toxic algae present are from the division Ochrophyta and to a lesser extent from the divisions Cryptophyta and Chlorophyta. Predicting estuarine algal blooms will enable proactive management regimes to be implemented, by forewarning managers of an impending algal bloom. To monitor algal blooms a telemetric buoy was deployed in the Berowra Estuary which records (at 15 minute intervals) Chlorophyll-a (CHLa), temperature, salinity and photosynthetically available radiation. Additional factors influencing algal dynamics considered in this study include: stratification, mixing, tidal flushing, flow rates, tidal excursion, rainfall, inorganic and organic substances, cell morphology, suspension, loss processes, competition, succession and specific growth rates. This study defines an “algal bloom” as the period when daily mean CHLa concentrations exceed the seasonal daily mean CHLa concentration for three or more consecutive days. An exceedance of the seasonal daily mean concentrations for less than three consecutive days, is termed a “perturbation”. Exponential CHLa alert thresholds were set at 0-4μg/L (low), 4-8 μg/L (moderate), 8-16μg/L (medium), 16-32μg/L (high), 32-64μg/L (very high) and >64μg/L (extreme). Data between 2004 and 2009 has been used to develop Artificial Neural Networks (ANNs) which predict daily mean, 10th and 90th percentile CHLa concentrations. The accuracy of the ANNs to predict CHLa concentrations decreased from one to three to seven days in advance respectively. The ANNs consist of a multi-layer perceptron architecture, with one hidden layer and use the Broyden-Fletcher-Goldfarb-Shanno training algorithm. These ANNs used the input variables, month, temperature, season, freshwater inflow, salinity recovery time (time taken for the estuary to return to mean salinity conditions following significant rainfall), tidal velocity, salinity, specific growth rates and observed CHLa data. Comparison of the variable contributions to the prediction performance of each ANN indicates that mean daily water temperature, tidal velocity and salinity recovery time were the key variables informing the predictions made by the ANNs.

Grantor dc:publisher

UNSW, Sydney

Year dc:date

2012

Author dc:creator

• Coad, Peter William

• Estuarine ecology
• Algal bloom
• Artificial neural network
• Water quality
• Enviornmental monitoring
• Environmental management

Statement dc:rights

• open access
• CC BY-NC-ND 3.0
• free_to_read

Licence

• https://purl.org/coar/access_right/c_abf2
• https://creativecommons.org/licenses/by-nc-nd/3.0/au/

Language dc:language

EN

Identifier

https://doi.org/10.26190/unsworks/15446

OAI identifier oai:identifier

oai:unsworks.library.unsw.edu.au:1959.4/51890