coping’with’mul./scale’and’balance’’ in’data’assimila · 2013-12-23 ·...
TRANSCRIPT
Coping with Mul.-‐Scale and Balance in Data Assimila.on
Kayo Ide UMD Jim McWilliams UCLA Zhijin Li NASA JPL
Yi Chao Remote Sensing SoluBons
USW12 for outer BC
ICC6
CenCoos.org
MoBvaBon I: MulB-‐Scaleness in the Model State § Geophysical dynamics exhibits mulB-‐scale (MS) phenomena à Modeling: MulB-‐ResoluBon (MR), MulB-‐Model (MM)
§ This work focuses on spaBally • MS/MR in horizontal
• Balance in horizontal and verBcal
USW12 ICC6
ICC1.5 ICC.75
!!!
x = xL + xS !![+ ...]xL:!!large+scale!
xS:!!smaller+scale!!
[Capet et al, 2008]
FormulaBon I. Standard 3DVar § Standard cost funcBon
• Background (xb,B) • ObservaBon (yo, R) with H: xày is the observaBon(forward) operator
§ Incremental cost funcBon
• Control variable: Δx = x – xb (departure from xb ) à x = xb+Δx • InnovaBon: d= yo–Hxb (diff between yo and xb in y-‐space)
SoluBon: Δxa = BHT (HBHT+R) -‐1 d − Single observaBon
!!!J(Δx)= 1
2ΔxTB−1Δx+ 1
2(HΔx−d)TR−1(HΔx−d)
!!!J(x− xb)= 1
2(x− xb)TB−1(x− xb)+ 1
2(yo −Hx)TR−1(yo −Hx)
!!!
Δxa =
B1lBllBNl
⎛
⎝
⎜⎜⎜⎜⎜⎜
⎞
⎠
⎟⎟⎟⎟⎟⎟
(Bll +Rll )−1(yl
o − xlb)
!!!⇐ p(x|y)=p(y|x)p(x)
p(y)
Contour: 1pt-‐correlaBon (�) Background: SST
SST [ICC6 ensemble]
.
Courtesy of T. Miyoshi
For Δxa to be MS, B must be MS
FormulaBon I. MulB-‐Scale Cost FuncBon for x § MulB-‐scale cost funcBon for Δx=ΔxL+ΔxS with B=BL+BS (AB: AddiBve B)
• InnovaBon: d= yo–Hxb [Wu et al, 2002] § Equivalent cost funcBons by decoupling the scales in control (Δx)
leading to the two independent cost funcBons
§ Successive (hierarchical) esBmaBon from larger to smaller scales i)
ii)
• Successive innovaBon* for ΔxS: d*= yo–H(xb+ΔxaL)
!!!J(Δx)= 1
2ΔxT (BL +BS )
−1Δx+ 12(HΔx−d)TR−1(HΔx−d)
!!!J(ΔxL ,ΔxS )=
12(ΔxL )
TBL−1ΔxL +
12(ΔxS )
TBS−1ΔxS +
12(HΔx−d)TR−1(HΔx−d)
!!!JS (ΔxS )=
12(ΔxS )
TBS−1ΔxS +
12(HΔxS −d)
T (R+HBLHT )−1(HΔxS −d)
!!!JL(ΔxL )=
12(ΔxL )
TBL−1ΔxL +
12(HΔxL −d)
T (R+HBSHT )−1(HΔxL −d)
!!!JS (ΔxS )=
12(ΔxS )
TBS−1ΔxS +
12(HΔxS −d
*)T (R*)−1(HΔxS −d*)
!!!JL(ΔxL )=
12(ΔxL )
TBL−1ΔxL +
12(HΔxL −d)
T (R+HBSHT )−1(HΔxL −d)
R*= R if ΔxaL=ΔxtL [Li et al 2013]
MoBvaBon II: Ocean ObservaBon Networks
h"p://www.sccoos.org/
Mooring network
u SpaBal distribuBon of observing networks is highly inhomogeneous, for example − Satellite images (SST) can be as high-‐resoluBon as the model state in horizontal at surface
− HR radar (surface velocity) can be highly concentrated and high-‐resoluBon
− Others can be extremely sparse
h"p://sccoos.org
SST on a good day HF radar network Glider network
FormulaBon II. MulB-‐Scale Cost FuncBon for x and y § ObservaBon classificaBon based on network resoluBon
§ MulB-‐scale decomposiBon for yD by smoothing
Leads to MS 3D-‐Var formulaBon by matching scales.
!!!y =
yD
yC
⎛
⎝⎜
⎞
⎠⎟ =
HD(xL + xS )
HC (xL + xS )
⎛
⎝⎜
⎞
⎠⎟ =
densecoarse
⎛
⎝⎜⎞
⎠⎟!!!!!!!!with!!
RD
RC
⎛
⎝⎜
⎞
⎠⎟
yo =yDo
yCo
⎛
⎝⎜
⎞
⎠⎟ =
yD.Lo + yD.S
o
yCo
⎛
⎝⎜
⎞
⎠⎟ ⇒ y =
yD.L
yD.S
yC
⎛
⎝
⎜⎜⎜
⎞
⎠
⎟⎟⎟=
HDxLHDxS
HC (xL + xS )
⎛
⎝
⎜⎜⎜
⎞
⎠
⎟⎟⎟$$with$$
RD.LL
RD.SS
RC
⎛
⎝
⎜⎜⎜
⎞
⎠
⎟⎟⎟
FormulaBon II. MulB-‐Scale Cost FuncBon for x and y § Successive ImplementaBon • EsBmaBon of dynamically important Large-‐scale first, then higher density to capture smaller scales
• Similar to Successive Covariance LocalizaBon (SCL: Zhang et al, 2009) à Use of B=BL+BS : AddiBve Background Covariance (AB 3D-‐Var)
• Explicit separaBon in y=yD.L+yD.S
§ MS 3D-‐Var: Easily extended to MS EnKF or 4D DA • Large-‐Scale: (LS)
• Small-‐scale: (SS)
!!!
JS (ΔxS )=12(ΔxS )
TBS−1ΔxS +
12(HDΔxS −dD.S
*)T (RD.S )−1(HDΔxS −dD.S
*)
!!!!!!!!!!!!!!!!!!!!!!!!!!!+ 12(HCΔxS −dC )
T (RC +HCBLHCT )−1(HCΔxS −dC )
dD.S* = dD.S −HDΔxL
a
!!!
JL(ΔxL )=12(ΔxL )
TBL−1ΔxL +
12(HDΔxL −dD.L )
T (RD.L )−1(HDΔxL −dD.L )
!!!!!!!!!!!!!!!!!!!!!!!!!!!+ 12(HCΔxL −dC )
T (RC +HCBSHCT )−1(HCΔxL −dC )
Simple DemonstraBon (1D-‐Var) § Experimental setup for MS/AB with {DL, DS} & SS with {DL, Dm1, Dm2, DS} • x: − xt is MS
− xb is MS − B may be MS/AB with (DL, DS)=(40, 5) & properly esBmated (σbL, σbS) may be SS with D =40, 20, 10, 5 & properly esBmated σb
• y=Hx may be MR
!!xnt = S0 ak
t cos(kπnN
+φnt ):!!!!!!!!!!!!!ak
t = k−γ !!!with!γ ⊂ [0,2]k=1
K
∑
patchy dense mixed: dense-‐coarse completely dense
!!xnb = S0 ak
b cos(kπnN
+φnb ):!!!!!!!!!!!!!ak
b = p0βk akt !!!with!βk ⊂U(0,1)
k=1
K
∑
SS (Single Scale)
Completely Dense Observing Network § Analysis (xa)
AB and MS work we with DS & DL
SS 3D-‐Var works OK at med. D
3DVar works OK at DS not at DL
§ Analysis Increment (Δxa)
AB and MS work Differently at LS , and hence at SS
MS captures LS more effecBvely than AB by sequenBal (successive) approach
Completely Dense Observing Network
3D-‐Var works beser at med. D than DS or DL
Black: target (xt-‐xb)
Patchy-‐Dense ObservaBon Network § Analysis (xa)
MS works much beser than AB
3DVAR works OK at med. D
3DVAR don’t work well: Beser at DS than DL
Patchy-‐Dense ObservaBon Network § Analysis Increment (Δxa)
AB and MS work differently at DS & DL
Standard 3D-‐Var don’t work well: Beser at DS than DL
MS captures LS more effecBvely than AB by sequenBal (successive) approach
Mixed ResoluBon Network § Analysis (xa)
MS works much beser than AB
3D work OK at medium D
3V work OK: Beser at DS than DL
Mixed ResoluBon Network § Analysis Increment (Δxa)
AB and MS work differently at DS & DL
3D don’t work well: Beser at DS than DL
MS captures LS more effecBvely than AB by sequenBal (successive) approach
Scale-‐Dependence: Analysis RMSE § Performance depends on treatment of MS in B (D) and H
H: patchy dense
H: mixed dense-‐coarse
H: completely dense
DL=35
D=20
D=10
DS = 5
MS
AB
smaller larger scale scale
MS AB DS=5
DL=35
D=10 D=20
California Coastal Ocean Data AssimilaBon System § Observing System Simula.on Experiments (OSSEs) • Model: Regional Ocean Modeling System (ROMS) − ResoluBon: 1km x 40 levels nested in low-‐resoluBon model − Atmos forcing: WRF at 2km
• Southern California Coastal Ocean Observing System (SCCOS)
− SST at 2km resoluBon − Surface (u,v) at 2km resoluBon − T/S profiling along 4 tracks at » 60km<DL separaBon between tracks » 10km<Ds, D3Dvar along track Up to 400m
− Balance (geostrophic & hydrosta9c) is incorporated
Bathymetry and OSSE T/S profiling posiBons
OSSE: RMSE Analysis Error in Time
§ Comparison between • NO DA • Standard 3D-‐Var (Ds in B) • MS 3D-‐Var (B & H) − Spin-‐up faster and
converges to smaller RMSE
(u,v) at z=30m SSH
T at z=30m S at 30m
OSSE: RMSE Analysis Error § VerBcal distribuBon of analysis RMSE At Day 3 (along-‐shore average)
NO DA
3D-‐Var
MS 3DVar
T S u v
California Coastal Ocean Data AssimilaBon System § Real Observa.on Experiments • IniBalizaBon: 01/01/2008 • Observing system (H)
• Performance: Comparison against independent data for bias − No DA − Standard 3D-‐Var − MS 3D-‐Var
h"p://www.sccoos.org/ h"p://sccoos.org
SST on a good day HF radar network
19
Glider network
Real ObservaBon: ValidaBon Against Independent Data § CALCOFI data sets
(o: not assimilated)
§ MS 3D-‐Var • ReducBon of bias • More observaBon is needed
Concluding Remarks § MS 3D-‐Var works well • Two main elements − Successive applicaBon of localizaBon from Large-‐scale to smaller scales − SeparaBon of observing system network
• ReducBon of bias
§ Extension to 4D & LETKF is straighvorward: • Scale separaBon Δx=ΔxL+ΔxS with ΔxL=ULwL ΔxS=USwS
• Scale dependent inflaBon
§ Coastal LETKF itself is challenging: maybe hybrid