Modeling dissolved oxygen for coastal monitoring and forecasting

The effectiveness of monitoring and forecasting dissolved oxygen (DO) levels in coastal regions is pivotal in the assessment of seawater quality, marine ecosystem activity, and aquaculture management. We propose a multi-stage model for coastal DO monitoring and forecasting, leveraging hourly-resolution data of seawater properties (e.g. water temperature, salinity, turbidity, pH, and velocity) collected using an Internet of Underwater Things (IoUT) sensor network. The sensors are located in the ”Smart Bay Santa Teresa”, northwestern Italy, near La Spezia. The measurement campaign started on March 2021 and is still ongoing in 2026. The collected data exhibit typical challenges of IoUT monitoring, such as power supply issues and loss of connectivity.

 

IoUT data are integrated with meteorological data provided by nearby stations (e.g. solar radiation, atmospheric pressure, air temperature, wind, rain), Copernicus Marine data referring to offshore conditions (including both blue and green seawater properties), and freshwater data from nearby rivers monitoring stations.

 

To reconstruct the missing data, we adopted separated regression models for the water temperature, salinity and oxygen. Each model is based on a residual deep learning approach using neural networks: the network is provided with an initial user-defined estimate, allowing the net to focus on unseen dynamics and unexpected behaviour. The adopted residual approach has demonstrated robustness in presence of large gaps in the data.

 

Once continuous monitoring is ensured, forecast DO levels over a horizon of a few days is performed. We currently focus on neural networks-based models, and tree-based regressors such as LightGBM. All these methods are benchmarked against baseline statistical models, such as Prophet and SARIMA. The tested models have shown encouraging ability to capture time-varying daily seasonal components, as well as extreme local events, which is of particular interest during peak blooms and hypoxia events.