Monitoring of harmful algal blooms (HABs) is crucial for maintaining both human and marine health. The Baltic Sea has seen an increase in the frequency and intensity of HABs in recent decades. This thesis investigates two applications of Gaussian Processes (GPs) for chlorophyll monitoring and forecasting.

First, satellite-measured chlorophyll concentration fields are augmented using in situ measurements collected by a front-tracking underwater autonomous vehicle (AUV), through a multi-step augmentation pipeline. The augmentation aims to reduce satellite measurement noise while keeping the broad spatial coverage. The results show that incorporating AUV measurements improves satellite measurements, demonstrating that in situ data can effectively complement satellite observations.

Second, two GP-based forecasting approaches are evaluated for predicting chlorophyll concentrations using temporal and spatio-temporal data: explicit kernel modeling and autoregressive modeling. The explicit kernel modeling approach requires explicit identification of data characteristics, which are encoded into the covariance function. The autoregressive approach learns dynamics implicitly from historical observations. Both methods show limited forecasting accuracy. The best-performing model is a purely temporal explicit kernel modeling approach, suggesting that this approach is more effective than learning dynamics implicitly.

Extending the models to the spatio-temporal domain produced mixed results. Explicit kernel modeling requires approximate inference for spatio-temporal data, resulting in reduced accuracy. Autoregressive models benefited from spatial information, but the forecasts failed to capture the chaotic behavior of algal blooms. Overall, the results highlight the limitations of GP based chlorophyll forecasting using chlorophyll observations alone and indicate that additional predictors are required for accurate forecasts.