• Speaker: Said OUALA, LAB STICC, IMT-Atlantique, Brest,France/ IMEDEA

Esporles, October 18, 2018.

Interpolating spatio-temporal dynamics given noisy observations with missing data is a key issue in environmental sciences. Among others, sea surface geophysical parameters are important drivers of oceanic and atmospheric circulation and the possibility of reconstructing highresolution fields is a challenge influencing several applications suchas tropical rainfalls forecasting and short term climate changeunderstanding.

Data assimilation based techniques are still the state-of-the-art approaches in the reconstruction of spatio-temporal geophysical fields. These methods heavily rely on an explicitly given dynamical model to compute several forward simulations. While model-driven representations are widely used to characterize the hidden dynamics present in our data the increasing availability of remote sensing observations and simulation datasets motivated the development of accurate and efficient data-driven models. Analog methods for instance are one of the first data-driven techniques to benefit from this deluge of observations and several techniques were developed to plug such data driven analog forecasting operators in a data assimilation scheme.

Initially introduced to solve classification and segmentation tasks, neural networks have shown very promising results in regression and identification problems. However, these powerful tools are barely used in the context of geophysical modelisation where deep non-linear characterization can be crucial to capture dependencies in spatio-temporal fields.

In our work, we investigate deep learning models to infer data-driven data assimilation dynamical priors using data. As a case-study, we consider satellite-derived Sea Surface Temperature time series off South Africa, which involves intense and complex upper ocean dynamics.

Date and Time: Monday, October 22, 12:30h

Place: IMEDEA Seminar Room

Source: IMEDEA (CSIC-UIB)