Data­-driven and learning-­based approaches for the spatio­temporal interpolation of SLA fields from current and future satellite­-derived altimeter data - IMT Atlantique Accéder directement au contenu
Communication Dans Un Congrès Année : 2018

Data­-driven and learning-­based approaches for the spatio­temporal interpolation of SLA fields from current and future satellite­-derived altimeter data

Résumé

The spatio­temporal interpolation of sea surface tracer fields from satellite­-derived altimeter data generally relies on model­-based schemes, the most popular ones being optimal interpolation schemes which exploit space­time covariance models. The ever increasing availability of in situ, remote sensing and simulation data make more and more appealing data­-driven alternatives, which may learn more complex representations of the underlying dynamics with a view to improving the reconstruction of fine­scale processes. We first review three categories of data-­driven schemes, namely patch­based super-­resolution models [Fablet et al., 2018], analog assimilation models [Lguensat et al., 2017] and neural­-network-­based assimilation models [Fablet et al., 2017]. We give more emphasis to the last two ones, which appear more generic. They are both stated within a Kalman­-based assimilation framework (namely the ensemble Kalman filter and smoother). They differ in the considered data­-driven dynamical model. The analog assimilation models exploit analog forecasting operators (especially locally-­linear analog operators) under the assumption that analog states share similar dynamics. By contrast, neural network (NN) architectures provide explicit representations of the dynamical operator. We focus on residual and convolutional architectures which may be interpreted as numerical integration schemes of differential equations (Fablet et al., 2017). For such data­driven and learning­-based schemes, the representativeness of the training data are critical issues. When dealing with high-­dimensional geophysical dynamics, the curse of dimensionality may make poorly relevant their straightforward application to the entire space­time domain of interest. We then discuss and introduce multiscale patch­-level representation as means to overcome these issues. We present numerical experiments for the spatio­temporal interpolation of SLA fields from satellite­-derived altimeter data using OSSE (Observing System Simulation Experiment) settings. We consider two types of satellite-­derived altimeter data, along-­track nadir data and upcoming wide­-swath SWOT mission. As case­ study region, we consider a region in the western Mediterranean sea with rich mesoscale and submesoscale dynamics. ROMS numerical simulations with a 0.02°x0.02° resolution over 5 years are used to implement the considered OSSE. The first 4 years are used for training. We apply the proposed interpolation schemes to the last one to evaluate their reconstruction performance. Overall, our results support a significant potential improvement for horizontal scales ranging from 20km to 100km with a gain of 42% (12%) in terms of SLA RMSE (correlation) with respect to the optimal Interpolation. Our results also suggest possible additional improvement from the joint assimilation of SWOT and along-­track nadir observations. We further discuss the pros and cons of data­-driven and learning-­based schemes for the reconstruction of SLA fields, especially future research directions to bridge model­-driven and data­-driven reconstruction schemes.
Fichier non déposé

Dates et versions

hal-01800511 , version 1 (19-07-2018)

Identifiants

  • HAL Id : hal-01800511 , version 1

Citer

Ronan Fablet, M Lopez-Radcenco, Said Ouala, Redouane Lguensat, L Gómez­navarro, et al.. Data­-driven and learning-­based approaches for the spatio­temporal interpolation of SLA fields from current and future satellite­-derived altimeter data. "25 Years of Progress in Radar Altimetry" Symposium, 2018, Punta Delgada, Azores, Portugal. ⟨hal-01800511⟩
367 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More