Predicting Pseudo-nitzschia harmful algal blooms along the Portuguese Coast using satellite-derived predictors

Research Question and Approach

This study addresses the challenge of predicting harmful algal blooms (HABs) caused by the diatom genus Pseudo-nitzschia along the northern Portuguese coast. Because these blooms can produce domoic acid—a neurotoxin that threatens shellfish safety and public health—there is a critical need for early-warning systems. The researchers developed a machine-learning framework that relies exclusively on satellite-derived data, such as sea surface temperature, upwelling indices, chlorophyll-a, and plankton functional types. To ensure the model reflects real-world conditions, the team implemented a strict spatio-temporal cross-validation strategy that withholds entire years and spatial clusters, preventing the model from "cheating" by learning from future or spatially adjacent data.

Key Findings

The researchers found that Pseudo-nitzschia blooms are moderately predictable using satellite data. Among the tested machine-learning models, ensemble tree-based methods performed best. A Random Forest model achieved a ROC-AUC of 0.74 using only environmental predictors, while an Extra Trees model improved this to 0.77 by incorporating biological variables. Feature-importance analysis revealed that seasonal timing, spatial location, and lagged environmental conditions (past states of the ocean) are the primary drivers of model decisions. Biological indicators, such as chlorophyll-a, serve to refine the bloom probability once the physical environment is already conducive to growth.

Traditional monitoring of HABs relies on labor-intensive in situ sampling, which is often limited by cost and logistics. By demonstrating that satellite-derived predictors can provide operationally relevant forecasts, this study offers a scalable tool for coastal management. The framework is specifically designed for eastern boundary upwelling systems, where wind-driven nutrient injection and riverine inputs create highly dynamic environments. This approach allows authorities to move beyond reactive monitoring toward proactive, data-driven risk assessments, potentially reducing the socioeconomic impact of shellfish harvesting closures.