Development of a deep learning-based feature stream network for forecasting riverine harmful algal blooms from a...

Area of Science:

• Environmental Science
• Water Resource Management
• Ecological Modeling

Background:

• Harmful algal blooms (HABs) are increasing globally, posing significant risks to water quality and ecosystem health.
• Effective management requires predictive models that capture the complex spatiotemporal dynamics of HABs and their drivers.
• Current models often struggle with multi-site forecasting and integrating diverse environmental factors within river networks.

Purpose of the Study:

• To develop and validate a novel Feature Stream Network (FSN) model for daily forecasting of cyanobacteria abundance across multiple sites in a river network.
• To quantify the spatiotemporal relationships between HABs and influencing factors, enhancing predictive accuracy and explainability.
• To assess the model's performance and utility for informing HAB management strategies.

Main Methods:

• A Feature Stream Network (FSN) model was designed, representing river sites as nodes in a directed acyclic graph to capture spatial connectivity.
• A segment-wise node connection structure and feature engineering-attention hybrid mechanism were employed to extract and transfer latent river segment features and handle temporal mismatches.
• The model was applied to the Nakdong River, South Korea, utilizing hydrological, environmental, and biological data for forecasting cyanobacteria abundance.

Main Results:

• The FSN model demonstrated high predictive accuracy for cyanobacteria abundance across multiple sites, with R-squared values ranging from 0.64 to 0.71.
• Feature importance analysis revealed that data from nearby monitoring sites significantly influenced forecasts, highlighting the model's ability to capture spatial hierarchy.
• Scenario analysis indicated that reducing total nitrogen loads and optimizing weir operations are effective strategies for HAB mitigation.

Conclusions:

• The FSN model offers an effective framework for multi-site, high-resolution forecasting of HABs in river networks, improving upon existing methods.
• The model provides valuable insights into the spatial dependencies and key drivers of HABs, aiding in targeted management interventions.
• The study underscores the potential of advanced modeling techniques for proactive water resource management and mitigating the impacts of harmful algal blooms.