Investigation of 4DVarNet Algorithm for Image Reconstruction of Suspended Particulate Matter Dynamics Data

High-resolution Suspended Particulate Matter (SPM) fields are vital for ecological-biogeochemical modeling, as they help study turbidity and micro-organism growth via photosynthesis. While satellite data provide valuable insights, cloud coverage often leads to missing information. Traditional interpolation methods, like DInEOF or eDInEOF, address this issue but can be computationally intensive. With advancements in computing power, Machine Learning (ML) and Deep Learning (DL) have emerged as efficient alternatives for reconstructing incomplete data. This project explores the application of a DL algorithm, 4DVarNet, to reconstruct SPM fields in the Dutch Wadden Sea. Based on 4D variational Data Assimilation (4DVar), 4DVarNet introduces two key innovations: 1) using Convolutional Neural Networks (CNN) to approximate dynamical systems, and 2) employing a neural iterative solver for image reconstruction. A variant of this algorithm, called Double LSTM, is also proposed. The performance of 4DVarNet will be compared to DInEOF and eDInEOF through an Observing System Simulation Experiment (OSSE). Here, 4DVarNet is trained on DELFT3D-FM model data and tested on satellite data (CMEMS), demonstrating its potential to improve efficiency and accuracy in reconstructing SPM fields.