Remote Sensing Remote sensing is a quasi-linear estimation problem Equation of radiative transfer: z TB (°K ) τ(z) T(z),α(z) TB (°K ) = TB e −τo o L z ∆ τ(z) = L ∫ α(z)dz z L + ∫ T(z)α(z)e 0 −τ(z) dz nepers m-1 temperature profile ∆ τo = τ(0) TB 0 o Lec22.5-1 3/27/01 D1 Radiation from Sky-Illuminated Reflective Surfaces TS TB z 1 L 2 4 0 TG terms 3 R=1-ε reflectivity TB (°K ) = RTs e −2 τo +εTG e sky temperature emissivity (specular surface) + Re L −τo −τo + 1 + 3 ground temperature L ∫ T(z)α( z)e − ∫0z α (z)dz dz 2 0 + ∫ T(z )α(z)e − ∫zL α (z)dz dz 4 0 L Lec22.5-2 3/27/01 ≅ ∫ T(z) ⎡α(z)e ⎣ 0 −τ(z) ⎤ dz for τo >> 1 ⎦ 4 D2 Temperature Weighting Function W(z,f,T) Terms: 1 , 2 , 3 For τo >> 1: 4 L TB (f ) ≅ To + ∫ T(z)W ( z,f,T(z)) dz 0 L Alternatively, TB (f ) ≅ To′ + ∫ ( T(z ) − To (z )) W ' ( z,f,To (z )) dz 0 Incremental weighting function: ∂TB W ' ( z,f,To (z)) = ∂T(z ) T (z ) o Note: we have ~linear relation: T(z) TB (f ) ↔ (not Fourier) Lec22.5-3 3/27/01 D3 Atmospheric Temperature T(z) Retrievals from Space α(ω) (single P) αo ∆ω ∝ pressure P area ∝ # molecules m 2 f ωo Therefore α o ≠ f(P) for P-broadening trace constituents L TB ≅ ∫ T(z) ⎡α(z)e ⎣ z −τ(z) ⎤ dz for τo >> 1 ⎦ 0 altitudes where intrinsic & doppler broadening dominate ω ≅ ωo ωo ± ∆ W(z,f) Lec22.5-4 3/27/01 pressure dominates ∆ω ωo ± 2∆ 0 α(z) D4 Atmospheric Temperature T(z) Retrievals from Space z altitudes where intrinsic & doppler broadening dominate ω ≅ ωo pressure dominates ∆ω ωo ± ∆ ωo ± 2∆ L −τ(z) ⎤ dz for τo >> 1 TB ≅ ∫ T(z ) ⎡d(z )e ⎣ ⎦ α(z) 0 W(z,f) 0 z α(ω) ∆ω increasing P ωo Lec22.5-5 3/27/01 ω f τ≅1 scale height W(f,z) D5 Atmospheric Temperature T(z) Retrievals from Below L ωo ωo ± ∆ W(f,z) ωo α(z) 0 W( f, z ) = α(z)e − ∫0L α (z)dz ~decaying exponentials, rate is fastest for ωo Temperature profile retrievals in semi-transparent solids or liquids where (1/α) >> λ: z transparency, If α(z) ≅ constant: frequency dependent L Lec22.5-6 3/27/01 0 W(f,z) D6 Atmospheric Composition Profiles L TB ≅ ∫ 0 α(z) − ∫zL α(z)dz ⎤ ⎡ dz ρ(z) • T( z)e ⎢⎣ ρ(z) ⎥⎦ W(z,f) L = TBo + ∫ [ρ(z) − ρo ( z)] Wρ′ ( z, f )dz 0 if viewed from space Because α(z) and W(z,f) are strong functions of the unknown ρ(z), this retrieval problem is quite non-linear and can be singular (e.g. if T(z) = constant). In this case, good statistics can be helpful. Incremental weighting functions defined relative to a nearly normal ρ(z) can help linearize the problem. Lec22.5-7 3/27/01 D7 Optimum Linear Estimates (Linear Regression) Parameter vector estimate pˆ = Dd “determination matrix” (d = [1,d ,..., d ]) ∆ 1 N data vector Choose D to minimize E ⎡(pˆ − p ) (pˆ − p )⎤ ⎣ Derive D : t { ∂ E ⎡(pˆ − p )t (pˆ − p )⎤ = 0 = ⎣ ⎦ ∂Dij ( t t t ∂ ⎡ = E d D −p ∂Dij ⎣ Therefore ⎦ i TH row of D ⎤ = E ⎡2d D j d − 2d p ⎤ − Dd p ( ) j i⎦ ) ⎦ ⎣ j DiE [dd j ] = E [pid j ] t t t t t ⎡ ⎤ ⎡ ⎤ DE dd = E pd ; Cd D = E ⎡dp ⎤ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ∆ Lec22.5-8 3/27/01 = Cd F1 Optimum Linear Estimates (Linear Regression) Therefore DiE [dd j ] = E [pid j ] t t t t t ⎡ ⎤ ⎡ ⎤ ⎡ DE dd = E pd ; Cd D = E dp ⎤ ⎣ ⎦ ⎣ ⎦ ⎣ ⎦ ∆ = Cd t The linear regression solution is pˆ = Dd where D = ⎡⎣Cd ⎤⎦ Lec22.5-9 3/27/01 −1 t⎤ ⎡ E dp ⎣ ⎦ F2 D is Least-Square-Error Optimum if: 1) Jointly Gaussian process (physics + instrument): pr (r ) = Λ m 1 ( 2π)N 2 Λ 1 2 ∆ ( − 1 ⎡(r −m) Λ ⎢ e 2⎣ −1 (r −m)⎤⎥ ⎦ ) t t = E ⎡(r − m) r − m ⎤ ⎣ ⎦ ∆ = E ⎡⎣r ⎤⎦, where r = r1,r2,...,rN 2) Problem is linear: data d = Mr + n + do parameter vector Lec22.5-10 3/27/01 noise (JGRVZM) F3 Examples of Linear Regression p̂ 1 − D : pˆ = [D11D12 ] ⎡1 ⎤ slope = D13 p̂ regression plane ⎢⎣d⎥⎦ D11 slope = D12 D11 0 slope = D12 d1 d d2 note change in slope Equivalently: p̂ = p + D12 ( d − d ) (E [•] ≡ • ) Lec22.5-11 3/27/01 F4 Regression Information 1) The instrument alone (via weighting functions) 2) “Uncovered information (to which the instrument is blind, but which is correlated with properties the instrument does see; D retrieves this too.) 3) “Hidden information (invisible in instrument and uncorrelated with visible information); it is lost. Lec22.5-12 3/27/01 G1 Nature of Instrument-Provided Information Assume linear physics: d = WT Where d = data vector [1,d1,d2,...,dN ] T = temperature profile [T1,T2 ,...,TM ] W = weighting function matrix W i = ith row of W Claim: N If T = ∑ ai W i and noise n = 0, i =1 then Tˆ = Dd = T, perfect retrieval (if W not singular ) Lec22.5-13 3/27/01 G2 Proof for Continuous Variables Claim: N If T = ∑ ai W i and noise n = 0, i =1 then Tˆ = Dd = T, perfect retrieval (if W not singular ) Let: W1(h) ∆ = b11φ1(h) ∆ W2 (h) = b21φ1(h) + b22φ2 (h) ∆ W3 (h) = b31φ1(n) + b32φ2 (h) + b33φ3 (h) ∞ ⎧0 Where: ∫ φi (h) • φ j (h)dh = δij = ⎨ ⎩1 0 ∆ φi;bij are known a priori (from physics). Then: d j = Lec22.5-14 3/27/01 ∞ ∫ T(h)W j (h)dh 0 G3 Proof for Continuous Variables ∞ ∆ Then: d j = ∆ ∫ T(h)Wj (h)dh W1(h) 0 W2 (h) = b21φ1(h) + b22φ2 (h) ∆ N If we force T(h) = ∑ k jWj (h) = b11φ1(h) ∆ i=1 Then: d j = N ∞ ∑ t ∫ (ki Wi (h)) Wj (h)dh ⇒ d = k W W = k Q t t t i=0 0 N∞ ⎛ i ⎞ ⎞⎛ N ∆ N = ∑ ∫ ki ⎜ ∑ bimφm (h) ⎟ ⎜ ∑ b jnφn (h) ⎟ ( dh) = ∑ kiQ ji ⎠ ⎠ ⎝ n =1 i=1 0 ⎝ m =1 i=1 Therefore So let Lec22.5-15 3/27/01 d = Qk where Q is a known square matrix −1 k̂ = Q d = k where Q is non-singular G4 Proof for Continuous Variables N Claim: If T = ∑ ai W i and noise n = 0, i =1 then Tˆ = Dd = T, perfect retrieval (if W not singular ) −1 k̂ = Q d = k where Q is non-singular So let Then: ∆ N ˆ T(h) = N ∑ kˆ iWi (h) = ∑ kiWi (h) = T(h) (exact) i=1 Equivalently: So: i =1 ( ) ( ∆ −1 T = Wk = W Q d = WQ D = WQ Q.E.D. −1 ) d = Dd −1 ∆ = "minimum information" solution Which is exact if T = Wk, n = 0 Lec22.5-16 3/27/01 G5 To what is an instrument blind? W1(h) ∆ = b11φ1(h) ∆ W2 (h) = b21φ1(h) + b22φ2 (h) An instrument is blind to T(h) components outside the space spanned by φ1,φ2,…,φN or, equivalently, by its W1, W2,…,Wn. By definition, the instrument is blind to any φj ⊥ Wi, for all i. Lec22.5-17 3/27/01 H1 Statistical Methods Can Reveal “Hidden” Components ∞ N In general, T(h) = ∑ k i Wi (h) + ∑ aiφi (h) =1 i 1 N + seen by N instrument channels all hidden components Extreme case: suppose φ1(h) always accompanied by 1 φN+1(h). 2 N N i =1 i =1 ˆ = ∑ k i Wi (h) = ∑ aiφi (h) Then our present solution: T(h) ( ) N Would become: T̂(h) = a1 φ1 + 1 φN+1 + ∑ aiφi 2 i= 2 shrinks with decorrelation Lec22.5-18 3/27/01 Thus hidden components can be “uncovered” if correlated with visible ones. H2 General Linear Estimate ∆ = β i −1⎤ ∞ T̂ = Dd where Di = ⎡WQ + ∑ aijφ j ⎢⎣ ⎥⎦i j =N+1 minimum “uncovered” information information Thus retrieval can be drawn only from the space spanned by N ⎛ ⎞ φ1, φ2 ,..., φN; β1,β2 ,...,βN ⎜ dimensionality is N; Tˆ = ∑ di Di ⎟ ⎝ ⎠ i=1 That is, N channels contribute N orthogonal basis functions to the minimum-information solution, plus N more basis functions which are orthogonal but correlated with the first N. Lec22.5-19 3/27/01 H3 General Linear Estimate As N increases, the fraction of the hidden space which is “uncovered” by statistics is therefore likely to increase, even as the hidden space shrinks. In general: a priori variance = observed + uncovered + variance lost due to noise and decorrelation. Example: 8 channels of AMSU versus 4 channels of MSU Lec22.5-20 3/27/01 AMSU and MSU are passive microwave spectrometers in earth orbit sounding atmospheric temperature profiles from above with ~10-km wide weighting functions peaking at altitudes from 3 to 26 km. Note the larger ratio of uncovered/lost power for AMSU. H4 Example: 8-Channel AMSU vs 4-Channel MSU MID-LATITUDES (TOTAL POWER = 1222 K2, 15 LEVELS) OBSERVED POWER UNCOVERED POWER LOST POWER MSU - 55° INCIDENCE ANGLE, LAND MSU - 0° ANGLE, NADIR AMSU - NADIR AMSU - NADIR 3 km WEIGTHING FUNCTION TROPICS (TOTAL POWER = 184 MSU – NADIR, LAND 10 6 12 22 26 16 (WINDOW) K2) UNCOVERED LOST AMSU - NADIR 0 Lec22.5-21 3/27/01 10 20 30 40 50 60 PERCENT POWER 70 80 90 100 H5