Top Banner
A Bayesian Methodology for Soil Parameters Retrieval from SAR Images M. Barber 1 , M. Piscitelli 2 , P. Perna 1 , C. Bruscantini 1 , F. Grings 1 , J. Jacobo-Berlles 3 , H. Karszenbaum 1 1 Grupo de Teledetección Cuantitativa, Instituto de Astronomía y Física del Espacio (IAFE), Buenos Aires, Argentina. 2 Facultad de Agronomía, Universidad Nacional del Centro de la Provincia de Buenos Aires (UNCPBA), Azul, Argentina. 3 Departamento de Computación, Facultad de Ciencias Exactas y naturales, Universidad de Buenos Aires (UBA), Buenos Aires, Argentina. 24-29 July, Vancouver, Canada. Session TU2.T03: Soil Moisture Remote Sensing II
16
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Page 1: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

A Bayesian Methodology for Soil Parameters Retrieval from SAR

Images

M. Barber1, M. Piscitelli2, P. Perna1, C. Bruscantini1, F. Grings1, J. Jacobo-Berlles3, H. Karszenbaum1

1 Grupo de Teledetección Cuantitativa, Instituto de Astronomía y Física del Espacio (IAFE), Buenos Aires, Argentina.

2 Facultad de Agronomía, Universidad Nacional del Centro de la Provincia de Buenos Aires (UNCPBA), Azul, Argentina.

3 Departamento de Computación, Facultad de Ciencias Exactas y naturales, Universidad de Buenos Aires (UBA), Buenos Aires,

Argentina.

24-29 July, Vancouver, Canada.

Session TU2.T03: Soil Moisture Remote Sensing II

Page 2: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

SAC-D

Launchedin 2011

SAOCOM

To beLaunched in 2015

Introduction and MotivationSurface soil moisture is a key variable related to:Climatic modeling (heat and mass transfer between Earth and

atmosphere).Hydrological applications (water dynamics and runoff, both closely

related with water erosion, risk of floods).Agronomical applications (vegetation and crops development).

Argentina is building an L-Band full polarimetric SAR system named SAOCOM:Offering the opportunity of monitoring surface soil moisture

content.In this context the goal of this work is to develop a soil moisture

retrieval algorithm.

As previous work a 3D laser profiler was developed and tested as part of the thesis done to complete my degree in Physics (Licenciatura en Física).

Page 3: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

SAC-D

Launchedin 2011

SAOCOM

To beLaunched in 2015

Concerning the speckle noise:• Arises due to the coherence nature of the imaging SAR system.• Leads to a grain-like appearance of SAR images.• Decreases the contrast and radiometric quality of SAR images.• It is usually reduced by spatial averaging at the expense of spatial

resolution.

Concerning the surface soil moisture:• It is a very heterogeneous variable across scales (high spatial and

temporal variability)

Limiting factors for retrieval of soil moisture from SAR backscattering coefficients

Concerning the forward models:• Non-negligible mismatches between model parameters and measured data

(statistical fluctuations, difficulty on parameterization of model variables, target’s heterogeneity)

Page 4: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

SAC-D

Launchedin 2011

SAOCOM

To beLaunched in 2015

Main features regarding Oh Model [Oh 2004]: Applies only over bare soils. Relates soil parameters

• volumetric soil moisture m (cm3/cm3),• normalized roughness rms height ks and correlation length kl (k=2p/l)

to polarimetric backscattering coefficients hh, vv, vh. Semi-empirical. Presented in an algebraic, explicit way. Constrained to the range 0.04<m<0.291 cm3/cm3 and 0.13<ks<6.98

(better agreement with ks<3.5). Straightforward inversion technique. Due to the insensitivity of the copolarized coefficients hh and vv to the correlation length kl and the problem of measuring it in the field, a rather desirable further formulation is made ignoring the correlation length so that:hh=f1(m,ks)vv=f2(m,ks)vh=f3(m,ks)

Available retrieval models: Oh Model

This simplified formulation of Oh Model will be used throughout this work.

Page 5: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

SAC-D

Launchedin 2011

SAOCOM

To beLaunched in 2015

Concerning Oh Model, we should be aware of:• It assumes a certain, unique m and ks value within a resolution cell

(homogeneuos surface’s target within a resolution cell)• It takes into account no speckle effects at all (i.e. achieved by averaging)These facts will implicitly be in any retrieval scheme that uses Oh model as the forward model.

Through a minimization procedure, Oh [Oh 2004] establishes an algorithm for retrieving soil moisture and roughness from a set of measured backscattering coefficients hh, vv and vh.

Shaded area encloses the Oh Model validity region

Page 6: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

SAC-D

Launchedin 2011

SAOCOM

To beLaunched in 2015

Some questions arise:At certain vh values, is it possible that a measured (hh,vv) pair would lie

outside the validity region? Yes, it certainly is (speckle noise and system uncertainities).

Which soil moisture estimate should be chosen in case of a measured (hh,vv) pair laying outside the validity region? A minimum-distance criterion could be an option but would yield to strongly noisy estimations.

In this sense a retrieval scheme based only in Oh Model is not suitable for real applications (i.e. those involving coherent imaging as from SAR systems).

Then, we should deal with the speckle noise affecting SAR images.

Page 7: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

SAC-D

Launchedin 2011

SAOCOM

To beLaunched in 2015

The Multiplicative Model includes terrain backscatter and speckle noise

Z=X.Yreturn

backscatterspeckle

How can we develop a model that includes terrain features as well as speckle noise in order to achieve a suitable retrieval for soil parameters, mainly for soil moisture?

Multiplicative Model for speckle

X and Y are assumed to be independent

If target is modeled without considering speckle, then Z=X.1

Page 8: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

SAC-D

Launchedin 2011

SAOCOM

To beLaunched in 2015

Bayes’ theorem

Providing the conditional density function the optimal unbiased estimators are:

LikelihoodPosterior

The Posterior is conditional probability of measuring m and ks given measured SAR backscattering coefficients hh, vv, vh.The Likelihood involves forward model as well as speckle model.In the Prior is included all a priori information about m and ks.The Evidence is an overall normalization factor.

Prior

Evidence

D={0.04<m<0.297 cm3/cm3,0.13<ks<3.5).

Page 9: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

SAC-D

Launchedin 2011

SAOCOM

To beLaunched in 2015

Model hypotheses

1) Xi=fi(M,KS) (i=1,2,3), so that fi represents the deterministic “typical” or average way in which the random variable Xi depends on the random variables M and KS (which represent target's m and ks).

2) It is sound to state E[Xi]=fi(E[M],E[KS]), (i=1,2,3).3) Speckle adds only a multiplicative noise of mean value E[Yi]=1,

(i=1,2,3).

Under these hypothesesE[Zi]=E[Xi]E[Yi]=fi(E[M],E[KS]), (i=1,2,3)

getting the proper average behavior for the returns in terms of the chosen forward model.

Z1=X1Y1

Z2=X2Y2

Z3=X3Y3

hh =f1(m,ks)vv =f2(m,ks)vh=f3(m,ks)

In our case the Oh Model is rewritten in terms of the multiplicative model as

Page 10: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

SAC-D

Launchedin 2011

SAOCOM

To beLaunched in 2015

In order to perform a numerical simulation, the following functions and model parameters are required:

1) Probability density function for M and KS.2) Prior.3) Number of looks n.

Model parameters

For the purpose of the simulation, it will be assumed that:

1) M and KS are normally distributed with sm=0.005 and sks=0.01.2) the Prior is uniformly distributed: U[0.01-0.35] for m and U[0.1-4.0]

for ks.3) n=3 and n=64 will be used for the number of looks.

Page 11: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

SAC-D

Launchedin 2011

SAOCOM

To beLaunched in 2015

Case n=3

Results

Units of m in cm3/cm3

One-sigma standard deviation contour lines

Estimated m

Page 12: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

SAC-D

Launchedin 2011

SAOCOM

To beLaunched in 2015

Results

Case n=64

Units of m in cm3/cm3

Estimated m

One-sigma standard deviation contour lines

Page 13: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

When using the Bayesian methodology, the retrieved soil moisture values cover the entire (hh,vv,vh)-space.

When significant multi-looking is present the Bayesian retrieval looks more compact around the corresponding contour lines of Oh Model indicating a correct asymptotical behavior.

In the same way, the one-sigma contour lines are also computed yielding an estimator for the retrieval errors.

The relative improvement due to the increasing of the number of looks is displayed by the one-sigma contour lines (error decreases from 0.07 cm3/cm3 to 0.03-0.04 cm3/cm3).

Results

Page 14: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

Concluding Remarks.

Soil parameters retrieval from SAR data presents two main sources of uncertainty: terrain heterogeneity and speckle noise. These issues were included into a retrieval scheme by using a Bayesian methodology.

Such a Bayesian methodology:• Needs only a forward model (no retrieval model

required) and a speckle model.• Gives an estimator of soil parameters as well as

their associated error.• Can include prior information in order to improve

the retrieval.

Page 15: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

Work in progress.

Performance assessment of the Bayesian model using a non-uniform (informative) prior.

Studying the sensitivity of the retrieval in terms of terrain heterogeneity parameters sm and sks.

Measuring in-situ roughness parameters over several soil tillages and soil types using the developed laser profiler (upper left figure) . This information will yield to a “roughness map” (upper right figure) which will link soil properties to probability distribution functions for s and will be further used as prior information.

Length [cm]

Width [cm]

Heig

ht

[cm

]

3D laser profiler output from a real soil.

Page 16: Barber_TU2.T03_hk_mb_casa_fg_hk_FINAL.pptx

THANK YOU