Advanced ocean modelling and machine learning to forecast Dinophysis acuminata blooms: A tool to support shellfish farming management

The presence of toxin-producers phytoplankton is a natural phenomenon that threatens marine ecosystems, endangers human health, and causes significant economic losses in shellfish harvesting. The Galician Rías Baixas (NW Spain) are one of the main mussels producing regions worldwide and the leading producer in Europe. Annual occurrence of Dinophysis acuminata , responsible for diarrhetic shellfish poisoning toxins, lead to a ban on the mussel harvesting for several months, each year. To help mitigate these impacts, this study explores the prediction of D. acuminata cells densities 3-days ahead in the outer and inner parts of three of the Rías Baixas (Arousa, Pontevedra and Vigo), testing three local machine learning (ML) models: Artificial Neural Network (ANN), Random Forest (RF) and Support Vector Machine (SVM). Local ML models were selected to account for the differences in occurrence and variability of D. acuminata densities across the different parts of each Ría. These ML models were assessed by (1) reducing the number of features through a feature selection approach to identify the most relevant ones, (2) exploring different sets of features and (3) comparing models trained with 7 and 30 days of past information. The dataset combined daily hydrodynamic and biological features, from 2013 to 2019, obtained from a high-resolution 3D hydrodynamic model (CROCO), and in-situ observations. Our results show that RF provided the best predictive performance. Increasing the number of days of past information did not significantly improve results, as similar averaged R 2 values were obtained for 7 and 30 days: 0.75 for Ría de Arousa, 0.72 for Ría de Pontevedra, and 0.67 for Ría de Vigo. Feature selection process showed that, as expected, previous cells densities of D. acuminata were essential for capturing bloom timing and amplitude. Also, the temperature, salinity, and the vertical and meridional components of current velocity were key predictors at outer stations of the Ría de Pontevedra and Vigo, where more features were required. In contrast, for the other stations, good predictions were achieved using only five features. This study represents one of the first attempts to predict D. acuminata in the Rías Baixas using local ML models. Our findings highlight the need for local approaches, as bloom dynamics vary between Rías and within different parts of each Ría. We also demonstrate the value of hydrodynamic model outputs to train ML models and compensate for the lack of long-term, spatially extensive in-situ data. • Output from a high-resolution 3D hydrodynamic configuration model, CROCO, was used as features for the prediction of D. acuminata. • To account for both spatial and temporal dependencies, predictions at a given station integrate spatially distributed data from neighboring locations and temporally lagged variables. • Random Forest outperformed Artificial Neural Network and Support Vector Machine models • A greater number of features were required to achieve accurate predictions in the outer areas of the Ría de Pontevedra and Vigo compared to their inner areas and to the whole Ría de Arousa.