Poster Session 4, Thursday, October 6, 11:00–12:40
Forecasting inorganic SPM concentrations using the 4DVarNet assimilation scheme applied to MODIS observations
The characterisation of inorganic (sediment) and detrital suspended particle dynamics provides information on turbidity, with applications to marine ecosystem studies and pollutant dispersion monitoring. However, due to natural and anthropogenic interconnected forcings, the dynamics of these particles remains difficult to understand and monitor. Numerical deterministic models still lack of capabilities in accounting for the variabilities observed by in situ and satellite sensors. The emergence of data-driven assimilation schemes is a possible relevant alternative, allowing parameter analysis and forecast, with possible increased capabilities in recovering fine-scale processes. Also, due to the large increase of both in situ and satellite measurements, more and more available data is coming. The long set of the two MODIS sensors’ observations is a good example for testing such particular aims. In that field of research, this work proposes a new forecasting approach based on a machine learning scheme applied to satellite observations only. This method is called 4DVarNet, which is a 4DVar scheme associated with two neural networks: a first one that represents the dynamical model, and a second one to perform the gradient descent optimisation. 4DVarNet is able to learn with time series of cloudy satellite images, contrary to other methods that typically needs cloud free L4 products as input data. The aim is to acquire as much as possible fine scale processes included in the L2 data, instead of using typically smooth L4 data. This work proposes to evaluate the quality of the predictions using our 4DVarNet architecture that learns on irregularly sampled data.
Fredéric Jourdin, Service Hydrographique et OCéanigraphique de la Marine, [email protected]
Ronan Fablet, IMT-Atlantique, [email protected]
Christophe Delacourt, CNRS UMR6538 LGO, [email protected]
Priscille Viellefon, CNRS UMR6538 LGO, [email protected]