The NW coast of the Iberian Peninsula is dominated by extensive shellfish farming, which places this region as a world leader in mussel production. Harmful algal blooms in the area frequent lead to lengthy harvesting closures threatening food security. This study developed a framework for the detection of Pseudo-nitzschia blooms in the Galician rias from satellite data (MERIS full-resolution images) and identified key variables that affect their abundance and toxicity. Two events of toxin-containing Pseudo-nitzschia were detected (up to 2.5 μg L⁻¹ pDA) in the area. This study suggests that even moderate densities of Pseudo-nitzschia in this area might indicate high toxin content. Empirical models for particulate domoic acid (pDA) were developed based on MERIS FR data. The resulting remote-sensing model, including MERIS bands centered around 510, 560, and 620 nm explain 73% of the pDA variance (R² = 0.73, p < 0.001). The results show that higher salinity values and lower Si(OH)₄/N ratios favour higher Pseudo-nitzschia spp. abundances. High pDA values seem to be associated with relatively high PO₄³, low NO₃⁻ concentrations, and low Si(OH)₄/N. While MERIS FR data and regionally specific algorithms can be useful for detecting Pseudo-nitzschia blooms, nutrient relationships are crucial for predicting the toxicity of these blooms.

This work presents new prediction models based on recent developments in machine learning methods, such as Random Forest (RF) and AdaBoost, and compares them with more classical approaches, i.e., support vector machines (SVMs) and neural networks (NNs). The models predict Pseudo-nitzschia spp. blooms in the Galician Rias Baixas. This work builds on a previous study by the authors (doi.org/10.1016/j.pocean.2014.03.003) but uses an extended database (from 2002 to 2012) and new algorithms. Our results show that RF and AdaBoost provide better prediction results compared to SVMs and NNs, as they show improved performance metrics and a better balance between sensitivity and specificity. Classical machine learning approaches show higher sensitivities, but at a cost of lower specificity and higher percentages of false alarms (lower precision). These results seem to indicate a greater adaptation of new algorithms (RF and AdaBoost) to unbalanced datasets. Our models could be operationally implemented to establish a short-term prediction system.

The Rias Baixas, located to the Northwest of Spain, along the northern boundary of the NW Africa upwelling system, is one of the most important areas of production of seafood in the world. In this area, thousands of commercial mussel platforms produced significant economic benefits for the region. Episodes of oceanic upwelling regulate the dynamics of movement of the Galician estuaries due to its North-South orientation and the summer winds of Northern component. Some species of Pseudo-nitzschia spp. are able to produce a toxic substance under specific environmental conditions. This substance is known as domoic acid (DA), and it is the responsible for the Amnesic Shellfish Poisoning (ASP), which affects several marine animals and climbs up in the food web causing severe human diseases. In order to avoid this toxicity, shellfish farms are closed temporally, forcing aquaculture and shellfish companies to interrupt its activity carrying out large economic losses. The increasing frequency and intensity of Proliferations harmful algal (HABs) in the area causes great economic losses so the development of systems of forecast should be of great interest for mussel production. The aim of this study is to expose the capability of physical oceanographic data to predict Pseudo-nitzschia spp. blooms and to link these different oceanographic conditions with chlorophyll a images and cell numbers of Pseudo-nitzschia spp.. Chlorophyll α images are obtained by means of color algorithms (processed with neural networks) using data from sensors of color as the MERIS (González et al., 2011), while the number of cells of Pseudo-nitzschia spp. are estimated using Support Vector Machines (SVMs) (González et al., 2014). The trade-off between these two methods (color algorithms and SVMs) results in an improvement of monitoring and management programs. The study was developed under the framework project Purga de Mar, which pretends to build up a prediction system for Harmful Algae Blooms allowing to take preventive measures in the Rias Baixas before toxicity appears. Artificial intelligence techniques are used to get this prediction, which are a net of sensors of environmental variables located in the study area, and data gathered using color sensor from satellite. These two methods (color algorithms and vector machines support) have been developed in order to evaluate the predictive power improved. The waters of the Rias Baixas are optically complex (case 2 waters), for what have been necessary color algorithms based on neural networks for the calculation of the chlorophyll α. using images from the MERIS sensor. Thus, 15 high resolution images (300 m) covering 670 Km×670 Km each over the study area were used. Cloudiness and availability were the reasons to use those ones. 228 chlorophyll measures were associated with 15 satellite images with a delay between in situ measures and satellite images of 30 minutes to 4 hours. The first step to create the algorithm is to apply an atmospheric correction and to mask the coast line, clouds, land and invalid pixels using the BEAM software. Two databases of in situ data were used to train neural networks. The first one, the INTECMAR database has chlorophyll a data recorded between 0 and 5 m depth from 2002 to 2004. The second one has chlorophyll α data recorded between 0 and 4 m depth during years 2006, 2007 and 2008. Fuzzy c-means (FCM) cluster algorithm was applied to MERIS reflectance to establish water clusters in order to determine the scope of the developed NNs. The number of clusters (3 clusters) was established as a function of the maximum degree of separation between them. Once the algorithm was developed, each data point was assigned to the cluster according to the maximum value for the membership function. Cluster#2 reflectance values were higher at lower wavelengths than cluster#1, while cluster#3 reflectance values were higher throughout the entire spectrum. Only one of the clusters, cluster #1, was representative and useful for training Neural Networks. Before estimating chlorophyll a with the algorithm, pixels of clusters#2 and 3 must be masked. These maps processed with color algorithms can be useful to study oceanographic processes which in turn could be linked to HABs events in the area. In the other side, to predict Pseudo-nitzschia spp. blooms with SVMs, 1471 registers of meteorological, water quality and cell count data were used. This information was compiled by INTECMAR between 1992 and 2000 in different stations located in the inner and in the outer part of the Rias. SVMs have been used as binary classifiers in presence/absence and bloom/ no bloom models. The dataset was divided in two sets: the training set (80% of data) and the test set (20% data). Training consists on a quadratic optimization method with constraints (using LIBSVM libraries). The model just needs two entrance parameters: cost (C) and ɣ of RBF Kernel (Gaussian radial basis function). Different models were created to select the most accurate one models of presence/absence were developed in order to distinguish samples with low abundance (<100cells/L); while models of Bloom/no Bloom were developed to identify blooms with high abundances (>〖10〗^5 cells/L). Another model B/NB was carried out with only presence samples. Models were developed using four combinations of variables. The model using all the variables showed the best result with overall accuracy value (78.57) and Kappa value (0.77) over the test set. Used variables were upwelling index in the last 4 days, temperature, salinity, previous week bloom, previous two week bloom, day of the year and Ria code. All these parameters were linear- scaled before being trained. There are two reasons to choose these parameters: their good fitting and the local predictive models available from the regional meteorological agency (Meteogalicia). Several studies have demonstrated the relationship between upwelling conditions and Pseudo-nitzschia spp. blooms. Through the study of images of chlorophyll a, phytoplankton blooms can be detected, and these can be related to ‘patches’ of Pseudo-nitzschia spp.. In order to assess the link between oceanographic conditions, chlorophyll maps, phytoplankton blooms and Pseudo-nitzschia spp. recounts, a study was carried out in the Rias Bias during the upwelling cycle in July 2008 (González et al., 2012). Color algorithms were applied to MERIS images in order to deliver chlorophyll maps. Surface temperatures were obtained from MODIS SST maps from the same dates that MERIS and wind measurements observed at a Seawatch buoy station sited in Cabo Silleiro were used to estimate upwelling index. MERIS images obtained during an upwelling event with concentrations up to 5mg/m^3 were mapped. SST maps showed temperature between 16-17ºC and temperature recorded by the buoy decreased about 1ºC. Similarly, another study with Pseudo-nitzschia spp. measurements was carried out in The Rias Baixas during 2007-2009 in order to assess the relationship between phytoplankton blooms, chlorophyll maps and Pseudo-nitzschia spp. ‘patches’ (Spyrakos et al., 2012). In this project MERIS maps show zones with high chlorophyll a levels where the highest Pseudo-nizschia spp. concentrations were recorded. This project manifests the importance of both tools (color algorithms and SVMs) to predict Pseudo-nitzschia spp. blooms. Estimation of chlorophyll linked to oceanographic features can be helpful in tracking ‘patches’ of Pseudo-nitzschia spp. since remote sensing is an important tool for monitoring, detecting phytoplankton blooms and their spatial distribution. SVMs have improved the capability to predict HABs since they use variables that can be operationally monitored in short or long term depending on the capability prediction (Meteogalicia or Puertos del Estado). Synergy between these two techniques has been proved to be a great tool in order to achieve an understanding of the dynamics of blooms and in order to get a better approach to bloom probability. Early detection systems have been proved to be the best tool in managing this kind of disasters. With the goal of improving Bloom prediction and detection, an oceanic modeling system was required to allow real-time forecasting Pseudo-nizschia spp. In order to improve forecasting systems include several lines of action in the future: the use of systems of automatic data located in the areas of production of mussels, the use of variables related to nutrients and data of detected domoic acid in the samples and the use of HF sensor data for a greater understanding of the surface dynamics of the estuary.