Loading…

Support Vector Machine-based method for predicting Pseudo-nitzschia spp. blooms in coastal waters (Galician rias, NW Spain)

•We analyzed 8years of Pseudo-nitzschia abundances and associated oceanographic data.•We developed SVM models for the prediction of Pseudo-nitzschia spp. blooms.•Results suggest a robust prediction model with great generalization power. Phytoplanktonic blooms in the coastal embayments (rias) at the...

Full description

Saved in:
Bibliographic Details
Published in:Progress in oceanography 2014-05, Vol.124, p.66-77
Main Authors: González Vilas, Luis, Spyrakos, Evangelos, Torres Palenzuela, Jesus M., Pazos, Yolanda
Format: Article
Language:English
Subjects:
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:•We analyzed 8years of Pseudo-nitzschia abundances and associated oceanographic data.•We developed SVM models for the prediction of Pseudo-nitzschia spp. blooms.•Results suggest a robust prediction model with great generalization power. Phytoplanktonic blooms in the coastal embayments (rias) at the NW part of Spain were mentioned for the first time in 1918 and since then they have been associated numerous times with negatives impacts to a very important economic activity in the area, mussel production. In this study, eight years of Pseudo-nitzschia spp. abundance and associated meteorological and oceanographic data were used to develop and validate support vector machine (SVM) models for the prediction of these diatoms. SVM were used to identify presence/below low detection limit, bloom/no bloom conditions of Pseudo-nitzschia spp. and finally to predict blooms due to these diatoms in the coastal systems of the Galician rias. The best SVM models were selected on the basis of C and γ parameters and their performance was evaluated in terms of accuracy and kappa statistics (κ). Regarding the presence/below low detection limit, bloom/no bloom models the best results in the validation dataset were achieved using all the variables: ria code, day of the year, temperature, salinity, upwelling indices and bloom occurrence in previous weeks. The best performing models were also tested in an independent dataset from the study area, where they showed high overall accuracy (78.53–82.18%), κ values (0.77–0.81) and true positive rates (62.60–78.18). In these models the bloom occurrence in previous weeks was identified as a key parameter to the prediction performance. In this paper, toxic Pseudo-nitzschia blooms could not be predicted due to limited information on toxin concentration and species composition. Nevertheless, this study demonstrates that the approach followed here is capable for high predictive performance which could be of great aid in the monitoring of algal blooms and offer valuable information to the local shellfish industry. The reliable prediction of categorical Pseudo-nitzschia abundances using variables that are operationally determined or short-term predicted could provide early warning of an impending bloom and could help to the development of strategies that could minimize the risks to human health and protect valuable economic resources.
ISSN:0079-6611
1873-4472
DOI:10.1016/j.pocean.2014.03.003