Forecasting of chlorophyll-a concentration using a wavelet-enhanced dynamic spatio-temporal graph convolutional network

Accurate prediction of chlorophyll-a (Chl-a) concentration is essential for harmful algal blooms (HABs) early warning, marine environmental monitoring, and coastal ecosystem management. Although deep learning-based models have made progress in Chl-a prediction, they often struggle to capture the dynamic spatial correlations of Chl-a and overlook the distribution shifts in its time series data. In this work, we propose a novel wavelet-enhanced dynamic spatio-temporal graph convolutional network (WDSTGCN) for Chl-a concentration prediction. This framework captures time-varying spatial correlations in marine environments and improves forecasting robustness under temporal distribution shifts. The model first disentangles complex Chl-a time series according to their frequency scales into two components: a low-frequency trend and high-frequency events, and applies separate spatio-temporal modeling to each component. Subsequently, the final Chl-a forecasts are generated through the effective fusion of low-frequency trends and high-frequency events. Furthermore, we design a dynamic graph generation module and integrate it into the spatio-temporal modeling stage to capture the time-varying spatial correlations of Chl-a. Extensive experiments on two real-world Chl-a datasets demonstrate that WDSTGCN outperforms strong baselines on both datasets. The results indicate that the proposed framework can better characterize complex Chl-a dynamics in coastal waters and provides a data-driven approach for multi-step Chl-a forecasting, with potential relevance to marine environmental monitoring.