Intro Inverse Modelling

Transcription

1 Intro Inverse Modelling Thomas Kaminski ( Thanks to: Simon Blessing (FastOpt), Ralf Giering (FastOpt), Nadine Gobron (JRC), Wolfgang Knorr (QUEST), Thomas Lavergne (Met.No), Bernard Pinty (JRC), Peter Rayner (LSCE), Marko Scholze (QUEST), Michael Voßbeck (FastOpt) 4th Earth Observation Summer School, Frascati, August 8 FastOpt

2 Overview BP: Basic concepts: from remote sensing measurements to surface albedo estimates () BP: Basic concepts: from remote sensing measurements to surface albedo estimates () TK: Introduction to inverse modelling TK: Intro tangent and adjoint code construction TK3: Demo: BRF inverse package and Carbon Cycle Data Assimilation System BP3: Monitoring land surfaces: applications of inverse packages of both surface BRF and albedo models FastOpt

3 References and Definitions A. Tarantola, Inverse Problem Theory and Methods for Model Parameter Estimation, SIAM, (987, 4) I. G. Enting, Inverse Problems in Atmospheric Constituent Transport, C.U.P. ()

4 Definitions Name Symbol Description Parameters p Quantities not changed by model, i.e. process parameters, boundary conditions State variables v(t) Quantities altered by model from time step to time step Control variables x Quantities exposed to optimisation, a combination of subsets of p and v(t = ) Observables o Measurable quantities Observation operator H Transforms v to o Model M Predicts o given p and v(t = ), includes H Data d Measured values of o

5 Statement of Problem Given a model M, a set of measurements d of some observables o, and prior information on some control variables x, produce an updated description of x. x may include parts of p and v(t = ) o = M( x)

6 Information and PDFs We seek true value but all input information is approximate Treat model, measurements and priors as PDFs describing distribution of true value Data and prior error easy Model error is difference between actual and perfect simulation for a given x Arises from errors in equations and errors in those parameters not included in x

7 Combining PDFs Operate in joint parameter and data space Estimates are combination of prior, measurement and model (black triangle) Estimate is multiplication of PDFs Only involves forward models Parameter estimate projection of PDF onto X-axis

8 Summary statistics from PDF Combination of PDFs (upper) and posterior PDF for parameters (lower) Can calculate all summary statistics from posterior PDF Maximum Likelihood Estimate (MLE) maximum of PDF

9 Tightening Model Constraint Note no viable solution If we cannot treat model as soft constraint we must inflate data uncertainty Model error hard to characterise

10 Gaussian Case G(x) = e (x µ) σ πσ Data: P d e (d ).6 Parameters: P x e x.6 Model: P M e (d x).6

11 Prior

12 Data

13 Prior plus Data

14 Model

15 Prior plus Data plus Model

16 Prior plus Data plus Model

17 PDFs and cost functions MLE most common quantity Maximise PDF numerically Most common P (x) e J so maximising P minimising J J usually called the cost (or misfit or objective) function Exponentials convenient because multiplying exponentials adding exponents

18 Perfect and imperfect models Imperfect Model Perfect Model Include observables ( o) in control variables Observables not included in control vector J = ( x x ) + ( o d)! (M( x) o) + J = ( x x ) + (M( x)! d) σx σobs σm σx σd x prior estimate for x, σ standard deviations. Allows explicit model error Different from weak constraint 4D-Var in which model state added to control variables If model imperfect, carried model error in σ d : σ d =σ obs + σ M This is the form usually used in data assimilation For linear M there is a closed solution Otherwise minimised by iterative algorithm, usually using gradients Can calculate posterior σ x by rearranging PDF

19 Covariances J = ( ( x x ) σ x + (M( x) d) ) σd Takes the more general form J = ( ( x x ) T C( x ) ( x x ) + (M( x) d) T C( d) (M( x) d) ) Off-diagonal terms represent correlation in uncertainty Persistent instrumental error Model error usually correlated Errors in initial conditions from previous forecast

20 Linear Gaussian case J = ( ( x x ) T C( x ) ( x x ) + (M x d) T C( d) (M x d) ) x opt = x + C( x )M T ( MC( x )M T + C( d)) ( d M x ) C( x) = C( x ) + M T C( d) M = d J(x opt ) dx C( x) does not depend on values, only uncertainties and model

21 How to solve it?. Minimisation Efficient minimisation algorithms use J(x) and the gradient of J(x) in an iterative procedure. Typically the prior value is used as starting point of the iteration. The gradient is helpful as it always points uphill. The adjoint is used to provide the gradient efficiently. Example: Newton algorithm for minimisation Gradient: g(x) = dj/dx(x) Hessian: H(x) = dg/dx(x) = d J/dx (x) At the minimum, x min : g(x min ) =, hence: g(x) = g(x) g(x min ) ~ H (x) (x-x min ) rearranging yields: (x min - x) ~ - H - (x) g(x) Figure: Tarantola (987) Smart gradient algorithms use an approximation of H(x) Figure: Fischer (996) FastOpt

22 How to solve it?. Error bars Hessian quantifies the curvature of the cost function Use Hessian at minimum to approximate C(x) - Figure taken from Tarantola (987) --> FastOpt

23 Exercise Do algebra for simple case Gaussian control variable observation Model: identity FastOpt