WWW.ABSTRACT.XLIBX.INFO
FREE ELECTRONIC LIBRARY - Abstract, dissertation, book
 
<< HOME
CONTACTS



Pages:   || 2 |

«Data assimilation methods for an oceanographic problem Didier Auroux1 and Jacques Blum2 Laboratoire J.A. Dieudonn´, Universit´ de Nice ...»

-- [ Page 1 ] --

Data assimilation methods for an

oceanographic problem

Didier Auroux1 and Jacques Blum2

Laboratoire J.A. Dieudonn´, Universit´ de Nice Sophia-Antipolis, Parc Valrose,

e e

F-06108 Nice Cedex 2, France. auroux@math.unice.fr

Laboratoire J.A. Dieudonn´, Universit´ de Nice Sophia-Antipolis, Parc Valrose,

e e

F-06108 Nice Cedex 2, France. jblum@math.unice.fr

1 Introduction

The dynamics of the oceans play a major role in the knowledge of our environment and especially in the Earth’s climate. Over the past twenty years, the new satellite techniques for observing the oceans, and especially the use of altimeter measurements, have greatly improved our knowledge of the oceans by allowing synoptic monitoring of the surface. The measurements of the seasurface height have clearly demonstrated the feasibility and the usefulness of satellite altimetry. It was with the availability of Topex/Poseidon data since 1992, that the oceanographic community began intensive exploitation of this new observational source. It has already given incomparable information to study the general circulation of the ocean, to estimate the energy levels of the upper ocean, and to examine the local dynamics of different regions of particular interest, such as the Gulf Stream area, the Kuroshio extension, the Antarctic circumpolar current and the tropical oceans.

At the interface between the two major components of oceanographic science, i.e. observations and models, lies the domain of so-called data assimilation (DA). DA covers all the mathematical and numerical techniques which allow us to blend as optimally as possible all the sources of information coming from theory, models and other types of data. Clearly these techniques may not only apply in oceanography but also to other environmental disciplines. DA allows us to recreate the time-space structure of a system from a set of information which has, in general, a large disparity in nature, in space-time distribution and in accuracy. There are two main categories of DA methods: variational methods based on the optimal control theory [Lio68] and statistical methods based on the theory of optimal statistical estimation. The prototype of the first class which is actually of interest here is the optimal control method which was first introduced in meteorology (see [Lew85], [LeD86], [Tal87]) and more recently for the ocean (see [Tha88], [She90], [Moo91], [Sch93], [Nec94], [Luo98]).

The prototype of statistical methods is the Kalman filter whose introduction 2 Didier Auroux and Jacques Blum in oceanography dates back roughly a decade (see, for example, [Ghi89] and [Ghi91]). The Kalman filter was extended to nonlinear cases ([Jaz70], [Gel74]) but it has been mostly applied in oceanography to quasi-linear situations of the tropical oceans ([Gou92], [Fuk95], [Fuk93], [Can96], [Ver99]). We also refer to the recent book of Bennett [Ben02] on inverse methods, both for oceanography and meteorology.

All DA techniques encounter major difficulties in practice for computing reasons: memory size and computing costs. The full Kalman filter would, in principle, require the manipulation of (N × N ) matrices where N is the state vector dimension which is typically 107 or 108 in an oceanic problem. The optimal control adjoint method often requires several hundred iterations of the minimization process to converge, thus implying an equivalent number of model runs.

In this paper, we first focus our interest on the use of the variational adjoint method in a relatively simple ocean model in order to try to reconstruct the four-dimensional ocean system from altimetric surface observations of the ocean. The variational method uses the strong constraint hypothesis, i.e. the ocean circulation model is assumed to be exact. The assimilation process is carried out by an identification of the initial state of the dynamical system which minimizes a cost function. This cost function is the mean-square difference between the observations and the corresponding model variables. The functional will be minimized using a numerical unconstrained optimization method such as the limited memory BFGS algorithm (see [Gil89]). The gradient vector is obtained analytically from the adjoint state, which can be interpreted as the Lagrange multiplier of the model equations. We then use a dual method, which consists in considering the model as a weak constraint.

The use of an observation vector as a Lagrange multiplier for this constraint allows us to consider the minimization problem in a dual way. The dual cost function, measuring the difference between the data and the model state corresponding to a vector of the observation space, is minimized in the observation space, still using the BFGS algorithm.

In section 2, we introduce the physical model used for the theorical and numerical results. The primal and dual methods applied to our ocean model are introduced in sections 3 and 4 respectively. Some numerical results are given in section 5. A few conclusions will be given in section 6.

2 Physical model

2.1 Quasi-geostrophy The system which governs the behaviour of the ocean is called the primitive equation system, constituted by the conservation laws of mass, momentum (Navier-Stokes equations), temperature and salinity. Most large-scale geophysical flows are based on the geostrophic equilibrium between the rotational Data assimilation methods for an oceanographic problem 3 effect due to the Coriolis force and the horizontal pressure gradient.





We will use here a simplified quasi-geostrophic ocean model. This model arises from the primitive equations, assuming first that the rotational effect (Coriolis force) is much stronger than the inertial effect. This can be quantified by the fact that the ratio between the characteristic time of the rotation of the Earth and the inertial time is small. This ratio is called the Rossby number.

The quasi-geostrophic model also assumes that the size of the ocean is small compared to the size of the Earth, and that this ratio is close to the Rossby number. Quasi-geostrophy finally assumes that the depth of the basin is small compared to its width (the ocean is supposed to be a thin layer of the Earth).

In the case of the Atlantic Ocean, all these assumptions are not valid, but it has been shown that this approximate model reproduces quite well the ocean circulations at intermediate latitudes, such as the Gulf Stream.

The thermodynamic effects are neglected, and we also assume that the forcing is due to the wind at the surface of the ocean and that the dissipation is essentially due to bottom and lateral friction.

2.2 Equations of the model

The ocean is supposed to be stratified in n layers, each of them having a constant fluid density [Hol78]. The quasi-geostrophic model is obtained by making a first order expansion of the Navier-Stokes equation with respect to the Rossby number [Ped79]. The model system is then composed of n coupled equations resulting from the conservation law of the potential vorticity. The

equations can be written as :

–  –  –

• n is the number of layers, Ψk is the stream function at layer k, Ψ is the vector (Ψ1,..., Ψn )T, • 4 Didier Auroux and Jacques Blum

• θk is the sum of the dynamical and thermal vorticity at layer k :

–  –  –

• and F1 is the forcing term, the wind stress applied to the ocean surface.

2.3 Boundary conditions The tridiagonal matrix W (used to couple the stream functions at different

layers) can be diagonalized :

–  –  –

at the bottom layer, in Ω×]0, T [, where

• Λ1,..., Λn is the adjoint vector, θk (Λ) = −∆Λk + (W T Λ)k is the vorticity corresponding to the adjoint T • state,

–  –  –

with sk = xk+1 − xk, ηk = J (xk+1 ) − J (xk ) and a ⊗ b : c → b, c a. The disadvantage of this formula is the need to store all pairs (sk, ηk ).

The L-BFGS algorithm ([Liu89]) is a limited memory version of the previous algorithm. Only the last M pairs are stored, M being often equal to 5. The

update formula is then :

–  –  –

4 Dual method

4.1 General description The primal method has many disadvantages. First, the minimization process is often stopped before convergence to the minimum, because of the size of the state vector. Moreover, it is also impossible to take into account a model error : in the previous section, we have supposed that the model and the equations were perfect. This is obviously not the case (for example, not all parameters are well known). The only solution to incorporate the model error 8 Didier Auroux and Jacques Blum into the minimization process is to add corrective terms to the model, consider them as part of the control vector, and add a third term to the cost function.

This is not computationally realistic because the size of the control vector would be multiplied by the number of time steps. Therefore, it is not possible to take into account in a straightforward way the model error in the primal variational approach.

A new approach to data assimilation problems has been recently introduced ([Amo95], [Ben92], [Cou97]). Rather than minimizing a cost function on the state space, the dual method consists in working in the observation space (which is smaller than the state space).

4.2 Dual algorithm

Instead of solving first the direct equations and then the adjoint equations in the primal variational approach, the dual method consists in solving first the adjoint equations in order to use the information contained in the observation vector, and then the direct equations in order to reconstruct a trajectory. The

dual algorithm for the quasi-geostrophic model can be constructed as follows :

• Let m be an observation vector that can be directly related to Ψ1 (assume that m is a vector containing an observation of a part of the ocean surface at different times ti ),

• Solve the adjoint equations (with a final condition equal to zero) :

–  –  –

5 Numerical results

5.1 Model parameters The numerical experiments are performed for a square three-layered ocean.

The basin has horizontal dimensions of 4000 km × 4000 km and its depth is 5 km. The layers’ depths are 300 meters for the surface layer, 700 meters for the intermediate layer, and 4000 meters for the bottom layer. The ocean is 10 Didier Auroux and Jacques Blum discretized by a Cartesian mesh of 200 × 200 × 3 grid zones. The time step is

1.5 hour. The initial conditions are chosen equal to zero for a six-year ocean spin-up phase, the final state of which being then the initial state for the data assimilation period. Then the assimilation period starts (time t = 0) with this initial condition (Ψk (0)), and lasts 5 days (time t = T ), i.e. 80 time steps. The numerical method used to minimize the cost functions is a limited memory BFGS quasi-Newton method. The M1QN3 code by Gilbert and Lemar´chal e ([Gil89]) is used for our experiments.

The experimental approach consists in performing twin experiments with simulated data. First, a reference experiment is run and the corresponding data are extracted. This reference trajectory will be further called the exact solution. Experimental surface data are supposed to be obtained on every fifth gridpoint of the model, with a time sampling of 7.5 hours (every 5 time steps).

Simulated surface data are then noised with a blank Gaussian distribution, and provided as observations for the cost function. The first guess of the assimilation experiments is chosen as the reference state of the ocean one year before the assimilation period. The results of the identification process are then compared to the reference experiment.

5.2 Exact solution, noised observations Fig. 1. Exact solution at the beginning (a), resp. the end (b), of the assimilation period.

Fig. 1 represents the stream function Ψ1 at the surface layer, at the beginning and at the end of the assimilation period. These fields will be useful to measure the identification of the initial state, and also the reconstruction of Data assimilation methods for an oceanographic problem 11 the stream function at the final time. One can observe the turbulent structure of the ocean, with a main current simulating a Gulf Stream type configuration.

Fig. 2. Noised extracted data at the surface layer (a) and corresponding state at the end of the assimilation period (b) The first part of fig. 2 represents the noised data extracted from the reference run, still at the surface layer. The second part of this figure is the corresponding state after a model run using the noised data as initial condition.

This experiment clearly shows the importance of data assimilation. The model will indeed not smooth the trajectory, and it is not possible to obtain good predictions by simply integrating the model with observation data as initial conditions.

5.3 Primal method

The initial estimated vector to start the minimization process is chosen to be the reference state of the ocean one year before the assimilation period.

The minimization process is stopped after 40 iterations, each iteration consisting of one integration of the forward direct model (in order to compute

J ) and one integration of the backward adjoint model (in order to compute J ). The result of the minimization is shown on fig. 3-a. The direct model is then integrated over the assimilation period, using the computed minimizer as initial condition, and the corresponding state of the ocean at the end of the assimilation period is shown on fig. 3-b.

We can notice that the stream function of the solution at time t = 0 at the surface layer is comparable to the exact solution at the same time, but to a 12 Didier Auroux and Jacques Blum Fig. 3. Result of the minimization of the primal cost function. Solution at the beginning (a) and the end (b) of the assimilation period lesser extent at time t = T. This can be explained by the fact that the primal algorithm gives more importance to the state at t = 0 than to any other time, as it is the control vector.



Pages:   || 2 |


Similar works:

«Langfristszenarien und Strategien für den Ausbau der Erneuerbaren Energien in Deutschland bei Berücksichtigung der Entwicklung in Europa und global FKZ 03MAP146 Arbeitsgemeinschaft Deutsches Zentrum für Luftund Raumfahrt (DLR), Stuttgart Institut für Technische Thermodynamik, Abt. Systemanalyse und Technikbewertung Fraunhofer Institut für Windenergie und Energiesystemtechnik (IWES), Kassel Ingenieurbüro für neue Energien (IFNE), Teltow Entwicklung der EEG-Vergütungen,...»

«MA/MFA Theatre Directing Postgraduate student coursebook 2015-16 Introduction Welcome Welcome to East 15 Acting School and thank you for choosing to study with us. Congratulations to all new students on gaining your place at East 15. We are delighted to welcome you to our Southend and Loughton campuses. And to all returning students – welcome back! East 15 offers an extraordinary dynamic portfolio of degrees that span many areas of training and study of performance for actors, directors and...»

«USGS Award No. G13AP00070 SITE RESPONSE MAPPING WITH ONE LESS PROXY: COLLABORATIVE RESEARCH WITH SDSU AND THE USGS FINAL TECHNICAL REPORT SUBMITTED: November 2014 Eric M. Thompson Geological Sciences San Diego State University 5500 Campanile Dr. San Diego, CA 92182-1020 ethompson@mail.sdsu.edu Tel: 858-848-0095 David J. Wald U.S. Geological Survey Golden, CO AWARD PERIOD: May 2013 to April 2014 Abstract The goal of this project is to process and analyze site-specific empirical amplification...»

«Proceedings of the 2012 Industrial and Systems Engineering Research Conference G. Lim and J.W. Herrmann, eds. Optical Container Code Recognition and its Impact on the Maritime Supply Chain Ravindra Shetty Agent technologies software Pvt ltd (www.agenttech.org) Bangalore, India Rebeca Cáceres Universidad Marítima Internacional de Panamá Panama City, Republic of Panama John Pastrana, Luis Rabelo Department of Industrial Engineering and Management Systems University of Central Florida, Orlando,...»

«1 © Copyright 2010 by the authors, Dr. Jolly Holden, and Dr. Philip J.-L. Westfall. All Rights Reserved. No part of this publication may be reproduced or transmitted in any form or by any other means, electronic or mechanical, including photocopy, microfilm, recording or otherwise, without the written permission from the authors. Previous revisions were published under the title An Instructional Media Selection Guide for Distance Learning, Copyright © 2005, 2006, 2007, 2008, 2009 Printed in...»

«Methoden zur Unterstützung bei der Entwicklung plattformübergreifender Benutzerschnittstellen Vom Fachbereich Informatik der Technischen Universität Darmstadt genehmigte Dissertation zur Erlangung des akademischen Grades eines Doktor-Ingenieurs (Dr.-Ing.) vorgelegt von Dipl.-Psych. Kai Richter geboren in Stuttgart Referenten der Arbeit: Prof. Dr.-Ing. J. L. Encarnação Prof. H. Wandke Tag der Einreichung: 16.02.2007 Tag der mündlichen Prüfung: 04.05.2007 D17 Darmstädter Dissertationen...»

«ZRÍNYI MIKLÓS UNIVERSITY OF DEFENCE Lajos Kossuth Military Science Course Military Science Doctoral School DÉZSI ZSOLT Introduction of Blended-Learning Method into Customs Officers’ Vocational Training Author Guide of PhD Dissertation Supervisor: Dr. Szelei Ildikó major reader Budapest, 2010 INTRODUCTION Nowadays, in parallel with technical development process, flare of training methods and tools and changes of requirements expected by employers, trainings based on information technology,...»

«Lehrstuhl für Landschaftsökologie der Technischen Universität München Scenarios assessing the viability of a lynx population in Germany. Szenarien für eine lebensfähige Luchspopulation in Deutschland. Stephanie A. Schadt Vollständiger Abdruck der von der Fakultät Wissenschaftszentrum Weihenstephan für Ernährung, Landnutzung und Umwelt der Technischen Universität München zur Erlangung des akademischen Grades eines Doktors der Naturwissenschaften (Dr. rer. nat.) genehmigten...»

«FAO ISSN 2070-7010 FISHERIES AND AQUACULTURE TECHNICAL PAPER Mud crab aquaculture A practical manual Cover photographs: Clockwise from top left: Scylla olivacea – dorsal view, courtesy of Queensland Museum; crablets of Scylla serrata, courtesy of Colin Shelley; larval rearing tanks covered with plastic to control aerosol contamination and assist in temperature control, courtesy of David Mann; earthen mud crab pond with netting around the pond, People’s Republic of China, courtesy of Chaoshu...»

«DER LANDKREIS OSTERHOLZ Die NBauO-Novelle 2012 Wesentliche Änderungen im Überblick Entwurfsverfasserveranstaltung 6. Juli 2012 DER LANDKREIS OSTERHOLZ Begrüßung Richard Eckermann Dezernent für Ordnung, Bauen und Umwelt Dominik Vinbruck Leiter des Bauordnungsamts Arend Bewernitz Leiter des Technischen Sachgebiets DER LANDKREIS OSTERHOLZ Inhalte NBauO 2012 im Überblick Bauanträge mit Servicegarantien Abschließende Hinweise und Fragen DER LANDKREIS OSTERHOLZ NBauO 2012 im Überblick DER...»

«Metadata and Memory: Lessons from the Canon of Memoria for the Design of Content Management Systems Author Stewart Whittemore Michigan State University Biosketch Stewart Whittemore is a PhD candidate in Rhetoric and Writing at Michigan State University. His dissertation research focuses on exploring the memory practices of knowledge workers. Abstract To date, most of the research on usability and content management systems has focused on the end-user products of such systems rather than on the...»

«Toward The Roles and Limitations Sustainability of Certification appenDiCeS June 2012 to accompany the Final Report of the State-of-Knowledge assessment of Standards and Certification Table of Contents Appendix A: Introduction to the Appendices............................................ A-1 Appendix B: Toward More Sustainable Production and Consumption, by Kai Lee and Ruth Norris.......................................»





 
<<  HOME   |    CONTACTS
2016 www.abstract.xlibx.info - Free e-library - Abstract, dissertation, book

Materials of this site are available for review, all rights belong to their respective owners.
If you do not agree with the fact that your material is placed on this site, please, email us, we will within 1-2 business days delete him.