Prediction of Sea Surface Chlorophyll-a Concentrations by Remote Sensing and Deep Learning

Accurate prediction of the spatiotemporal distribution of chlorophyll-a (Chl-a) is essential for evaluating marine ecosystem health and predicting ecological disasters. Current methods struggle to capture short-term variability and periodic trends in Chl-a, especially in noise-prone coastal regions. This study aims to enhance the prediction of marine Chl-a concentrations by introducing the chlorophyll-a concentration prediction model (ChlaPM), which was developed on the basis of a convolutional long short-term memory (ConvLSTM) network. The model integrates recent spatiotemporal feature extraction (RSTFE), periodic feature extraction (PFE), and denoising fusion (DNF) modules to effectively capture short-term spatiotemporal changes and periodic variations in Chl-a concentrations. In this study, the performance of ChlaPM in single-step and multistep predictions was evaluated using monthly average Chl-a remote sensing data spanning 1998–2023. The results indicate that compared with the RSTFE model, the ChlaPM model achieves substantial reductions in the root mean square error (RMSE) of 53.84%, 53.58%, and 49.70% for predicting Chl-a concentrations 1 month, 3 months, and 6 months into the future, respectively. These findings highlight the effectiveness of ChlaPM in addressing short-term variability and periodic trends and significantly enhances the accuracy of Chl-a prediction. Future work will focus on integrating additional relevant marine variables into the prediction model to further improve its prediction capabilities.