Internship in geoscience and artificial intelligence

Background of internship:
The assessment of underwater sighting range is required in marine operations, offshore engineering, coastal management and leisure activities. This transparency is mainly a function of the amount, size and nature of the microscopic particles in suspension within the water column. At the ocean surface, these particles of organic and mineral origin induce slight variations in water colour that can be detected from space. Deeper in the ocean, at basin to coastal scales, the fate of these particles can be modelled with hydrodynamic and biochemical equations. Usual techniques for merging information that comes from satellite data and models are referred as “model-driven” assimilation methods, as they mainly rely on model equations. But these methods often require complex schemes; also they are computationally expensive and are sensitive to nonlinearities. New, upcoming methods referred as “data-driven” assimilations are emerging and make use of machine learning and artificial intelligence and benefit from the rising amount of data and model simulations available in the oceanographic community.

Purpose of internship:
In the field of geophysics and artificial intelligence, SHOM cooperates with two laboratories located near Brest (Lab-STICC and LGO) to develop data-driven approaches for the processing of time series of images acquired by oceanographic satellites where large gaps in data occur due to cloud cover.
The aim of this internship is to test these methods on satellite images of the ocean colour. The study will focus on the Bay of Biscay and the English Channel. For this analysis, a database of more than 15 years of acquired satellite images of the ocean colour in this region is available. The trainee will apply these methods on a partial set of data and will assess the accuracy of results using remaining data. The predictive capabilities of the forecasting of water transparency will also be studied. Depending on the progress, additional information from model simulations, also available at SHOM, could be included in the process.