Data-driven assimilation:

I am currently working on an alternative to the classical model-driven data assimilation. The proposed technique is completely data-driven and is based on a catalog of historical data. The basic idea is to combine analog forecasting and classical ensemble data assimilation. Presentation of the method is given in this chapter of the book Machine Learning and Data Mining Approaches to Climate Science. This book grew out of the workshop on Climate Informatics held in Boulder, Colorado. The Python code for the Analog Data Assimilation (AnDA) is available at:

Sketch of the forecast step in stochastic data assimilation schemes using pure (top) and analog (bottom) dynamical models. We consider the xy-plane of the Lorenz-63 chaotic model. We track five statistical members with the variability depicted by ellipsoids accounting for the covariance structure.

Uncertainty quantification in data assimilation:

The Ensemble Kalman Filter is a widely used data assimilation method. This algorithm is a stochastic approach based on Monte Carlo simulations and very useful to account for nonlinear dynamics and high dimensional problems. In this article of the Quarterly Journal of the Royal Meteorological Society, I proposed a statistical algorithm (Expectation-Maximization) to estimate the error covariance matrices in such nonlinear state-space models used in EnKF. I also presented an application for the estimation of physical parameters for subgrid-scale processes. The Python code for the Covariance Estimation in Data Assimilation (CEDA) is available at:


Estimation of model error covariance matrices (model Q and observation R) using an Expectation-Maximization (EM) algorithm coupled with the Extended Kalman Smoother (EKS) or the Ensemble Kalman Smoother (EnKS).

Satellite synergy:

Another part of my research is dedicated to the synergy in ocean remote sensing, studying the statistical relationships between SST and SSH mesoscale fields. In the Agulhas current, we identified 4 hidden relationships corresponding to different responses of geostrophic currents to surface temperature local variations. Results are published in the journal Transactions on Geoscience and Remote Sensing. Contact me (pierre.tandeo@imt-atlantique) if you want to get the Matlab version of the latent class regression.


Animation of the affectation probability maps (blue 0 to red 1) for 4 latent ocean surface dynamics. The results are obtained using a latent regression model between daily interpolated microwave SST (Sea Surface Temperature) and SSH (Sea Surface Height) fields over the year 2004 within the Agulhas region.

Then, using the 4 dynamical modes identified by the latent class regression, I generated surface currents from single SST snapshots. Results are presented in this proceeding of the ICIP conference (10% best paper award).


Comparison between the AVISO surface currents (b) and different predicted currents from a single SST image using latent class regression (c), linear regression (d) and nonlinear regression (e).

I am also interested in the synergy between ocean colour, sea surface temperature and altimetry in the Malvinas current. See this presentation at the COSPAR meeting for more details.

Synergy between geostrophic surface currents and SST from AMSRE (left) and Chlorophylle-a from MODIS (right) in the Brazil-Malvinas convergence zone.

Spatio-temporal variability and satellite interpolation:

During my Ph.D. at IFREMER, I was working on the interpolation of satellite data using statistical approaches. First, I learned the SST temporal variability using infrared METOP-AVHRR data. Results are published in the journal Stochastic Environmental Research and Risk Assessment.


Estimation of the SST (Sea Surface Temperature) temporal correlation (left), the SST variance of the stationary distribution (middle) and the measurement error variance of the SST satellite sensor METOP. The parameters are estimated using one-year time-series of SST anomalies (SST METOP – SST Reynolds) at each spatial location.

Then using geostatistical tools on the same satellite data, I learned the SST spatial variability using anisotropic structures represented as ellipsoids to model the spatial correlation lengths and angles. Results are published in the journal Remote Sensing of Environment.


High resolution SST (Sea Surface Temperature) from the METOP-AVHRR sensor (left) and the corresponding estimated spatial anisotropic variability represented as ellipsoids (right).

Finally, using the spatial and temporal correlation lengths, we apply a Kalman filter and smoother to interpolate raw SST satellite data (with missing values due to clouds). Results are published in the Copernicus Marine Environment Monitoring Service (at, SST_NWS_SST_L4_REP_OBSERVATIONS_010_023 product).


Results of the daily spatio-temporal interpolation of 0.02 degree x 0.02 degree Pathfinder AVHRR SST data from 1982 to 2012 in the Atlantic European north west shelf seas.

Correction of atmospheric conditions:

Another part of my work is dedicated to the correction of SST data using drifting buoys. The idea is to fit a statistical model between delta SST (difference between satellite and in situ) and the atmospheric conditions. Results for the AATSR sensor onboard Envisat are published in the journal Geoscience and Remote Sensing Letters.