Improved prediction of chlorophyll-a concentrations using advancing graph neural network variants

Area of Science:

• Environmental Science
• Data Science
• Water Resource Management

Background:

• Accurate estimation of harmful algal blooms (HABs) is crucial for surface water protection.
• Chlorophyll-a (Chl-a) is a key indicator for algal concentration but is influenced by complex environmental factors.
• Existing methods struggle to integrate diverse data sources across various scales for long-term Chl-a simulation.

Purpose of the Study:

• To develop a deep learning (DL) framework for long-term Chl-a simulation.
• To effectively integrate irregularly measured water quality data and regularly measured climate data.
• To leverage graph neural networks (GNNs) for modeling spatiotemporal dependencies in Chl-a prediction.

Main Methods:

• A DL framework with two blocks: one for water quality observations and another for climate data.
• Benchmarking GNN architectures (ChebNet, GCN) for climate data encoding and modeling water quality stations as graph nodes.
• Utilizing a gating mechanism to integrate outputs from both data processing blocks.

Main Results:

• The proposed GNN-based framework significantly outperforms baseline models, achieving up to 47% improvement in R².
• The combination of Graph Convolutional Network (GCN) and Long Short-Term Memory (LSTM) demonstrated superior performance.
• The gating mechanism further enhanced prediction accuracy, improving R² by 12%.

Conclusions:

• The proposed GNN-variant framework offers a robust machine learning approach for Chl-a prediction.
• This method effectively aggregates spatiotemporal information for reliable water quality monitoring.
• The framework shows promise for advancing the accurate estimation of harmful algal blooms.