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

Area of Science:

• Environmental Science
• Water Quality Monitoring
• Data Science

Background:

• Estuarine dissolved oxygen (DO) monitoring faces challenges due to short-term fluctuations, multiscale temporal variability, and observation noise.
• Limited continuous monitoring data and site-dependency complicate robust DO prediction.
• Extracting temporal patterns and inter-variable dependencies from routine data is crucial for effective DO forecasting.

Purpose of the Study:

• To develop an effective method for high-frequency dissolved oxygen forecasting in estuarine environments.
• To address challenges of noisy data, multiscale temporal variability, and complex variable couplings.
• To propose RimeSG-STGNN, an adaptive denoising multiscale graph neural network for improved DO prediction.

Main Methods:

• Proposed RimeSG-STGNN framework combining SG-based noise reduction, RIME-driven parameter optimization, multiscale temporal feature extraction, and hybrid graph learning.
• Utilized 15-minute monitoring data from four estuarine sites in the United States, incorporating eight water quality variables.
• Employed graph neural networks to jointly model noisy temporal patterns and inter-variable dependencies.

Main Results:

• RimeSG-STGNN demonstrated superior prediction accuracy compared to baseline models across four estuarine datasets.
• Achieved significant RMSE reductions of approximately 66% and 73% at specific US estuarine sites.
• Validated the model's effectiveness for high-frequency DO forecasting in complex estuarine monitoring scenarios.

Conclusions:

• RimeSG-STGNN is an effective tool for robust dissolved oxygen forecasting in high-frequency estuarine monitoring.
• The proposed framework successfully addresses data noise and complex temporal dependencies.
• Further research will explore the model's cross-region applicability using larger, multi-region datasets.