The use of data driven deep learning models to predict and monitor Harmful Algal Blooms (HABs) has evolved over the years due to increasing technologies, availability of high frequency data, and statistical prowess. Despite the prowess of these data driven models, they are limited by inherent model structure and uncertainty in the generating process. To overcome the limitations of data driven models, in this research, we introduced the concept of data assimilation (DA) to account for model errors and incorporate new observations into the data driven deep learning HABs prediction model. Data assimilation is a computational method that enhances the precision of predictions in dynamic systems by combining real-time observations with model forecasts. In this study, we developed 100 Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU) to make one-day ahead prediction of chlorophyll-a, an indicator of HABs, using high-frequency pH, temperature, specific conductivity, turbidity, dissolved oxygen, saturated dissolved oxygen, and oxidation-reduction potential (ORP) data. We used an Ensemble Kalman Filter (EnKF) approach to assimilate chlorophyll-a observations of greater confidence into the HABs prediction model. We explored different assimilation frequencies to observe the appropriate timesteps required for introducing new information into the modeling system. The results showed improved chlorophyll-a prediction, as forecasted by the system when DA is applied. We found that increasing assimilation frequency tends to provide improved chlorophyll-a prediction, with daily assimilation having RMSE of 0.03 μg/l for GRU and 0.02 μg/l for LSTM, while monthly assimilation resulted in RMSE of 3.63 μg/l for GRU and 3.59 μg/l for LSTM. The study revealed the potential application of DA strategy to enhance the accuracy and reliability of deep learning models for HABs monitoring. In the presence of new chlorophyll-a observations, findings from this research inform on the appropriate frequency to which such information can be incorporated into a HABs prediction model framework. This process ensures that the model provides timely and accurate predictions to support effective HABs management and decision-making efforts.

Effective management of water resources entails the understanding of spatiotemporal changes in hydrologic fluxes with variation in land use, especially with a growing trend of urbanization, agricultural lands and non-stationarity of climate. This study explores the use of satellite-based Land Use Land Cover (LULC) data while simultaneously correcting potential evapotranspiration (PET) input with Leaf Area Index (LAI) to increase the performance of a physically distributed hydrologic model. The mesoscale hydrologic model (mHM) was selected for this purpose due to its unique features. Since LAI input informs the model about vegetation dynamics, we incorporated the LAI based PET correction option together with multi-year LULC data. The Globcover land cover data was selected for the single land cover cases, and hybrid of CORINE (coordination of information on the environment) and MODIS (Moderate Resolution Imaging Spectroradiometer) land cover datasets were chosen for the cases with multiple land cover datasets. These two datasets complement each other since MODIS has no separate forest class but more frequent (yearly) observations than CORINE. Calibration period spans from 1990 to 2006 and corresponding NSE (Nash-Sutcliffe Efficiency) values varies between 0.23 and 0.42, while the validation period spans from 2007 to 2010 and corresponding NSE values are between 0.13 and 0.39. The results revealed that the best performance is obtained when multiple land cover datasets are provided to the model and LAI data is used to correct PET, instead of default aspect-based PET correction in mHM. This study suggests that to minimize errors due to parameter uncertainties in physically distributed hydrologic models, adequate information can be supplied to the model with care taken to avoid over-parameterizing the model.

This study explores the use of satellite-based LULC (Land Use / Land Cover) data while simultaneously correcting potential evapotranspiration (PET) input with Leaf Area Index (LAI) to increase the performance of a physically distributed hydrologic model. The mesoscale hydrologic model (mHM) was selected for this purpose due to its unique features. Since LAI input informs the model about vegetation dynamics, we incorporated the LAI based PET correction option together with multi-year LULC data. The Globcover land cover data was selected for the single land cover cases, and hybrid of CORINE (coordination of information on the environment) and MODIS (Moderate Resolution Imaging Spectroradiometer) land cover datasets were chosen for the cases with multiple land cover datasets. These two datasets complement each other since MODIS has no separate forest class but more frequent (yearly) observations than CORINE. Calibration period spans from 1990 to 2006 and corresponding NSE (Nash-Sutcliffe Efficiency) values varies between 0.23 and 0.42, while the validation period spans from 2007 to 2010 and corresponding NSE values are between 0.13 and 0.39. The results revealed that the best performance is obtained when multiple land cover datasets are provided to the model and LAI data is used to correct PET, instead of default aspect-based PET correction in mHM. This study suggests that to minimize errors due to parameter uncertainties in physically distributed hydrologic models, adequate information can be supplied to the model with care taken to avoid over-parameterizing the model.