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

Accurate estimation of harmful algal blooms is essential for protecting surface water. Chlorophyll-a (Chl-a), commonly used as a proxy for estimating algal concentration, is influenced by a broad range of weather and physicochemical factors that operate across various spatial and temporal scales. This study aims to propose a deep learning (DL)-based framework for long-term Chl-a simulation, consisting of two separate blocks for processing multi-modal sources together: one for incorporating irregularly measured water quality observations and the other for integrating climate data measured at constant time steps. Besides a fully connected network for encoding irregular water quality observations, we benchmark several state-of-the-art graph neural network (GNN) architectures, including ChebNet and Graph Convolutional Network (GCN), for encoding continuous climate data. Specifically, we represent water quality stations as nodes in a graph, model the spatiotemporal dependencies between these nodes, and utilize the learned relationships to predict Chl-a simulations simultaneously across all nodes in the graph. Additionally, we introduce a gating mechanism to integrate the outputs from the two blocks. The performance of advanced GNN models is evaluated using a daily dataset from the upper Han River basins in South Korea. The results indicate that our proposed models are promising, outperforming several baseline models developed for similar objectives with improvements up to 47 % in the R². In particular, the combination of the GCN algorithm with Long Short-Term Memory (LSTM) in our DL framework achieves superior performance. We then conduct further analyses to assess the effectiveness of the gating mechanism, revealing that it enhances prediction performance by achieving a 12 % improvement in the R² compared to the model without the gating mechanism. We conclude that the proposed GNN-variant framework shows promise as a robust machine learning-based approach for aggregating spatiotemporal information to achieve reliable Chl-a predictions.