Polarimetric SAR tomography of tropical forests using P-Band TropiSAR data

Transcription

1 Polarimetric SAR tomography of tropical forests using P-Band TropiSAR data Yue Huang, Laurent Ferro-Famil, Cedric Lardeux SAPHIR team, IETR University of Rennes 1, France 19/01/2011

2 Objectives Global objectives: Characterization of tropical forests Tree height & biomass Robustness assessment First year objectives: Implement classical tomographic approaches Single-pol (SP) and full-pol (FP) configurations Estimate tree height Second year: Estimate biomass or biomass-related quantities (extinction) Compare to POL-inSAR Test baseline configurations Work done up to now POLTOM data obtained at the end of August 2010 Classical tomographic approaches implemented, SP and FP versions Very fast tree height estimation: hybrid HR methods Tries at L band Comparison with POL-inSAR under investigation

3 Basics Distributed Scatterer: Unconditional Model Tree height and ground topography estimation Tomography basics Acquisition geometry MB-PolInSAR: Polarimetric tomography Localize scatterers in z direction & extract their physical features Nominal resolution δ h 1 L tomo, ambiguity height h 1 ds

4 Basics Distributed Scatterer: Unconditional Model Tree height and ground topography estimation Tomography basics General signal model y = A(θ)x+n C m x C d : source (reflected) signals (d elements) m d steering matrix A(θ) = [a(θ 1 ),...,a(θ d )] m-element steering vector, a(θ i ) = a(z i ) = [1, exp{jk z2 z i }, exp{jk zm z i }] T Focusing Continuous spectral estimators: Matched filter (Fourier), Capon... Discrete ones: estimate the d sources

5 Basics Distributed Scatterer: Unconditional Model Tree height and ground topography estimation Power spectrum of various environments Natural environment (h-distributed scatterers): Continuous Spectrum Objects (h-localized scatterers): Discrete Spectrum Objects embedded in natural environment: Mixed-Spectrum

6 Basics Distributed Scatterer: Unconditional Model Tree height and ground topography estimation Classical spectral estimators for tomographic imaging Nonparametric spectral estimators Continuous spectrum Beamforming: ẑ = arg max{a H Ra} 1 a H R 1 a } a: steering vector R: sampled data covariance matrix. Capon: ẑ = arg max{ Moderate resolution Parametric spectral estimators Discrete Spectrum E.g. Weighted Signal Subspace Fitting (WSF) ẑ = arg min tr{p A ÊsWÊH s } P A : orthogonal projection matrix of steering matrix A. W: weighting matrix E s : signal subspace High resolution Lack of adaptation to the type of spectrum!

7 Basics Distributed Scatterer: Unconditional Model Tree height and ground topography estimation Unconditional MB-PolinSAR signal model y u = d σi x ui a(z i )+n i=1 Valid for Distributed Scatterers with speckle affected responses (Ground) Stochastic source signal (white for each observation ) x i = σ i x ui with x ui N m (0, C i ) C i describes interferometric coherence y u N m (0, R y ) σ i, z i estimated from R y

8 Basics Distributed Scatterer: Unconditional Model Tree height and ground topography estimation Conditional MB-PolinSAR signal model y c = d σi x ci a(z i )+n i=1 Valid for coherent scatterers: (Double Bounce) x i = σ i x ci is deterministic (frozen) over N observations (looks) y c N m (Ax,σnI) 2 σ i, z i estimated from Ax

9 Basics Distributed Scatterer: Unconditional Model Tree height and ground topography estimation Hybrid SAR signal model Mixture of coherent and distributed scattering contributions * (Sauer et al: 2007) y = y c + y u = d1 i=1 σi x ci a(z i ) + d2 i=1 σi x ui a(z i ) + n Demonstration: Two-component [ hybrid source ] signal x(l) [ = [x u (l), ] x c (l)] σu 0 σc σ with R xu = and R 0 σ xc = c u σ c σ c Source covariance matrix R x = R xu + R xc = (σ u + σ c ) [ 1 ρhd ρ hd 1 ] Source correlation ρ hd = 1 1+UCR with UCR = σ u : uncoherent to coherent intensity ratio. σc ρ hd 1, R x singular Hybrid model correlated unconditional model ρ hd correlation of UMs.

10 Basics Distributed Scatterer: Unconditional Model Tree height and ground topography estimation Model adaptive WSF spectral estimator Conventional estimators MUSIC: ẑ = arg min{a H E n E H n a} uncorrelated scatterers Det-ML: ẑ = arg max tr{a(a H A) 1 A HˆRy } coherent scatterers Stoch-ML: ẑ = arg min P A ˆR y P A + ˆσ 2 P A noncoherent scatterers E n : noise subspace; P A : projection matrix to the signal subspace; P A : orthogonal projection matrix. Model adaptive estimator: Weighted Subspace Fitting ẑ,ˆt = arg min E s W 1/2 AT 2 F

11 Basics Distributed Scatterer: Unconditional Model Tree height and ground topography estimation Natural Environment: Hybrid tomographic method Principle Hybrid: estimate both continuous and discrete spectral components Fast : Simple estimators, CAPON (canopy), WSF (ground and volume mass center) CAPON Backscattered power spectrum P(z) WSF (order=2) Ground topography z g Phase center of the volume z v Tree top height: z top = {z P(z) = P(z v )-3dB} Easy extension of the proposed tomographic approach to the fully polarimetric case.

12 Paracou Nouragues Test site: Paracou, French Guiana TropiSAR Campaign ONERA SETHI (P band) 6 tracks Resolution Azimuth: δ a = 1.245m Range: δ r = 1m Tomographic resolution δ z 20m

13 Paracou Nouragues Discrimination of bare soils and forested areas (MOS) SP MOS FP MOS

14 Paracou Nouragues Forest profile estimation Hybrid approach: Tree Top: black line Ground: gray line HH HV VV FP No profile post-processing (peaky aspect) Very Yue similar Huang, Laurent FP and Ferro-Famil, co-pol Cedricresults Lardeux POLTOMSAR TROPISAR presentation 19/01/2011

15 Paracou Nouragues Ground topography and tree height estimation Shadowed regions should be masked out

16 Paracou Nouragues ROI validation for P9 and P10 (logged plots)

17 Paracou Nouragues ROI validation for P11 (Undisturbed forest) and P12 (logged plots)

18 Paracou Nouragues Full-pol results Worse than HH case, need for more sophisticated approaches parametric tomography (2nd year)

19 Paracou Nouragues First results on Nouragues HH VV HV No profile post-processing (peaky aspect) Well Yue handled Huang, Laurent strong Ferro-Famil, topography Cedric Lardeux (specific POLTOMSAR processing) TROPISAR presentation 19/01/2011

20 Paracou Nouragues First results on Nouragues HH VV HV No profile post-processing (peaky aspect) Well Yue handled Huang, Laurent strong Ferro-Famil, topography Cedric Lardeux (specific POLTOMSAR processing) TROPISAR presentation 19/01/2011

21 Paracou Nouragues Conclusions Tomographic results Excellent data quality at P band L band images for TOMSAR -Pol-inSAR? to be checked SP and FP tomograms provide very good profile description Topography can be handled (specific processing) Fast approach to estimate tree height and ground topo Short term developments Tomogram filtering (peaky aspect) & shadow masking Systematic comparison to ground measurements (Paracou & Nouragues) Mid term developments Full-pol parametric tomographic focusing (height and extinction) L band data processing & comparison to POL-inSAR