Reconstruction of missing satellite turbidity data of the Southern North Sea using a neural network.
Supervisors: Alexander Barth, Aida Alvera (Uni. Liege)
Gap-filling of satellite imagery is key to improving the resolution and quality of data for various ocean variables, the presence of clouds remains as one of the key issues in remote sensing and the neural networks have proven to be an effective method to overcome this matter. The present work was conducted to reconstruct turbidity fields in a study area along the Belgian coast, known for high cloud cover and optically complex waters, for the year 2020. The reconstructions were performed using the DAta Interpolating Convolutional AutoEncoder (DINCAE), a neural network with the structure of a convolutional autoencoder. Among the main results, a product of high temporal and spatial quality was obtained, which was able to evaluate the turbidity trends and the seasonal patterns in the study area, and even to generalise and approach the quality of a set of data used for cross-validation. The best reconstruction achieved an RMSE of 0.26 in log FNU, compared to another reconstruction method as Data Interpolating Empirical Orthogonal Functions (DINEOF), which achieved a RMSE of 0.31. The spatial variability, assessed by variogram analysis, showed that DINCAE can preserve the original variability of the data. Our results allow us to confirm that using a multi-source product with high spatial and temporal resolution can be used in DINCAE with positive generalisations according to the hyperparameters chosen. Lastly, further analysis suggests that the addition of other variables is suited for performing a robust multivariate reconstruction in the neural network and it will increase the level of detail of diverse field outputs.