Prediction of Temporal Dissolved Oxygen Concentrations in a Lake Using Remote Sensing and Machine Learning

Dissolved oxygen (DO) levels are crucial for aquatic life, especially under climate change, making continuous monitoring essential for effective lake management. However, local measurements are often costly and time-intensive, whether collected through field campaigns or permanent gauges. This study investigates the feasibility of using remote sensing techniques, coupled with machine learning; to track and estimate DO in a shallow eutrophic lake. Because DO cannot be directly measured with optical sensors, we first identify optically sensitive parameters—chlorophyll-a (Chl-a), temperature, and water depth—that correlate statistically with ground-measured DO. A two-step pipeline is then introduced: (1) estimating water level changes, Chl-a, and surface temperature from satellite data, and (2) predicting DO based on these derived parameters.

 

Model development starts with developing three separate models to estimate Chl-a (Sentinel-2), water level changes (Sentinel-1), and lake surface temperature (MODIS), using the Google Earth Engine Python API for data processing and analysis. Subsequently, both remotely sensed parameters and local measurements are used to train a DO prediction model. The training procedure explores 16 machine learning frameworks with hyperparameter tuning, using a 70%–15%–15% time-series split for training, validation, and testing, implemented in scikit-learn and Optuna. Search stopped with the model with R² values of 0.89 and 0.64 and mean absolute errors of 0.81 mg/L and 1.29 mg/L for locally measured and predicted test datasets, respectively. These results highlight the potential of combining remote sensing-derived parameters with machine learning to estimate DO, an otherwise non-optically measurable parameter.

 

This approach offers a cost-effective alternative for modeling continuous temporal variations in DO and supports comprehensive temporal assessments of DO concentrations in shallow eutrophic lakes. Ultimately, this framework shows promise for broader applications and generalizations, thereby contributing to the effective monitoring of non-optical water quality parameters and advancing sustainable aquatic ecosystem management.