Remote sensing is a source of a great quantity of observations of the earth system, and espacially of the ocean surface. Satellite images are a efficient way to monitor on a large spatial and temporal scale some physical and biological processes.
The illustration shown here presents a map of concentration of a secondary pigment of phytoplankton (fucoxanthin). In this example the pigment is a key factor of biodiversity in the ocean as it can be linked to some phytoplakton functionnal types. Nevertheless, this pigment is not directly measured by the satellite, but the optical signal measuread by the satellite (the so-called luminance) was associated with pigment concentration using a classification algorithm that combined satellite and in-situ data.
This example shows that remote sensing observations are naturally linked to algorithmic and methodological problems.
A major challenge of data sciences and machine learning apply to environmental sciences is the following question: what can be learned from the data itself ? The Earth surface (including the oceans) are more and more observed. The dynamic of this system is defined by physical laws for which with have only a partial knowledge (due to a lack of knowledge in spatial and temporal resolution, parametrization or physical processes). Could we fill partially these gap by applying a learning aglorithm on the data themself
In the illustration shown, we perform a forecast of the total suspended matter on a small region (7x7 pixels). This evolution is mainly driven by wind and tide forcing. In our case, we only feed a deep neural architecture with sequence of images. The neural network is able to reproduce tide cycle and variability due to the wind effect without been given any physical assumption. It shows that a part of the ocean dynamic can be extracted from the data themselves.
Data Assimilation is a scientific domain that consists in combining data from several sources : a numerical model and some observations. It gives a theorical and pratical framework to estimate the state of a system (e.g. velocities in the ocean) that maximizes the likelihood of this state given the observations and a numerical model. The bayesian framework gives also a natural framework for uncertainty quantification.
The illustration shown gives an example of an application of data assimilation to the merging of two sources of data : satellite derived velocities of the surface ocean and successive positions of drifting buoys.