Author(s): Cruz, Rafaela C. ; Reis Costa, Pedro ; Krippahl, Ludwig ; Lopes, Marta B.
Date: 2022
Persistent ID: http://hdl.handle.net/10400.1/18877
Origin: Sapientia - Universidade do Algarve
Subject(s): Time series; Forecasting; Artificial Neural Networks; Biotoxins; Shellfish contamination; Harmful algal blooms
Harmful algal blooms (HABs) and the consequent contamination of shellfish are complex processes depending on several biotic and abiotic variables, turning prediction of shellfish contamination into a challenging task. Not only the information of interest is dispersed among multiple sources, but also the complex temporal relationships between the time-series variables require advanced machine methods to model such relationships. In this study, multiple time-series variables measured in Portuguese shellfish production areas were used to forecast shellfish contamination by diarrhetic she-llfish poisoning (DSP) toxins one to four weeks in advance. These time series included DSP con-centration in mussels (Mytilus galloprovincialis), toxic phytoplankton cell counts, meteorological, and remotely sensed oceanographic variables. Several data pre-processing and feature engineering methods were tested, as well as multiple autoregressive and artificial neural network (ANN) models. The best results regarding the mean absolute error of prediction were obtained for a bivariate long short-term memory (LSTM) neural network based on biotoxin and toxic phytoplankton measurements, with higher accuracy for short-term forecasting horizons. When evaluating all ANNs model ability to predict the contamination state (below or above the regulatory limit for contamination) and changes to this state, multilayer perceptrons (MLP) and convolutional neural networks (CNN) yielded improved predictive performance on a case-by-case basis. These results show the possibility of extracting relevant information from time-series data from multiple sources which are predictive of DSP contamination in mussels, therefore placing ANNs as good candidate models to assist the production sector in anticipating harvesting interdictions and mitigating economic losses.(c) 2022 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
