Harmful algal blooms pose a significant threat to marine ecosystems, aquaculture industries, and human health. To mitigate these risks, agencies around the globe perform regular monitoring and operate early warning systems based on expected risk levels. However, bloom dynamics can be influenced by a large range of physical and biological factors, leading to high uncertainty in predictions of future blooms. Here, we explore the effectiveness of ensemble models for forecasting risk of algal blooms and associated toxins in Scotland, employing a diverse set of candidate models, including tree-based approaches, neural networks, and hierarchical Bayesian regression. These models predicted the probability that algal densities or biotoxin concentrations would exceed a threshold (either 'amber' status in the traffic light guidance system, or 'detection') in the next week using publicly available environmental products combined with regulatory monitoring data from dozens of locations in Scotland (2015-2022
Alexandrium spp., Dinophysis spp., Karenia mikimotoi, Pseudo-nitzschia spp. [+ delicatissima, serriata groups], domoic acid (DA), okadaic acid / dinophysistoxins / pectenotoxins (DSTs), paralytic shellfish toxins (PSTs)). The forecasted probabilities from the candidate models were used as inputs for a stacking ensemble model. Compared to individual candidate and null models, the ensemble models consistently improved forecasting performance across two years of withheld out-of-sample validation data, as assessed by five distinct performance metrics (ensemble skill scores among metrics and targets: mean = 0.499, middle 95 % = 0.214-0.900
skill score give improvement over the null model, with 1 indicating perfect performance). Performance varied by monitoring target, with best forecasts for DSTs (mean ensemble skill: 0.747) and poorest for K. mikimotoi (mean ensemble skill: 0.334). Autoregressive terms and regional spatiotemporal patterns emerged as the most informative predictors, with effects of environmental conditions contingent on the algal density or toxin concentration. Our results demonstrate the clear advantage of the ensemble approach. The operational implementation of these models provides probabilistic forecasts to enhance Scotland's monitoring program and early warning system. Ensemble modelling leverages the combined strengths of the wide array of modern techniques available, offering a promising path toward improved forecasts.