A self-adaptive denoising and multi-scale spatio-temporal graph network for dissolved oxygen prediction

Dissolved oxygen (DO) in estuarine environments often exhibits strong short-term fluctuations, pronounced multiscale temporal variability, observation noise, and complex coupling among multiple variables, while the variables that can be continuously monitored over long periods are often limited and site-dependent. Therefore, how to effectively extract temporal patterns and inter-variable dependency structures from routinely monitored variables has become a critical issue for achieving robust DO prediction. To address these issues, this study proposes RimeSG-STGNN, an adaptive denoising multiscale graph neural network for high-frequency DO forecasting. The framework combines SG-based noise reduction, RIME-driven parameter optimization, multiscale temporal feature extraction, and hybrid graph learning to jointly model noisy temporal patterns and variable dependencies. Experiments were conducted using 15 min monitoring data from four estuarine sites in the United States, with eight water quality variables as inputs. Compared with representative baseline models, RimeSG-STGNN consistently achieved superior prediction accuracy across the four datasets, with RMSE reductions of approximately 66% and 73% at the Research Creek station in North Carolina and the Potters Cove station in Narragansett Bay, United States, respectively. These results demonstrate the effectiveness of RimeSG-STGNN for DO forecasting in high-frequency estuarine monitoring scenarios, while its broader cross-region applicability will be further examined in future studies using larger multi-region datasets.