pres_h2a_okconvert.ppt

Dual Polarised Entropy/alpha Decomposition and Coherence Optimisation for Improved Forest Height Mapping

Z-S Zhou, P. Caccetta, E. Lehmann, A. Held – CSIRO, AU

S. McNeill – Landcare, NZ

A. Mitchell, A. Milne and I. Tapley - CRC for Spatial Information & UNSW

K. Lowell - CRC for Spatial Information & University of Melbourne

CSIRO: Dual Polarised Entropy/alpha Decomposition and Coherence Optimisation for Improved Forest Height Mapping

Contents

Introduction

Dual Polarisation Entropy/alpha Decomposition

Partial Polarised Coherence Optimisation

Joint Processing of Dual Polarised Entropy/alpha Decomposition and Coherence Optimisation for Forest Height Mapping

Conclusions and Future Work

* Acknowledgments: The Australian Department of Climate Change and Energy Efficiency, Forestry Tasmania, Geoscience Australia, JAXA.


IntroductionNCAS (National Carbon Accounting System)

--Land Cover Change Project

• Australia’s National Forest Cover Change Program

[Department of Climate Change (AGO)]

• Continental Landsat Archive 1972-2006,2007,2008, 2009, 2010, 2011….

• Forest Change Products [methods]

• Forest Vegetation Trends

www.climatechange.gov.au/ncas

http://www.climatechange.gov.au/ncas


– Consistently processed time series entire Australian continent

– 19 time periods used so far

– Few clouds !?

NCAS Landcover Change Project: Optical Time Series Data (25m)

1972

2010

Mosaic ~ 400 scenes

……

….

…

…

…


Forest-Non-Forest : Digital Classification of Cloud Free Image at one time Epoch

NCAS 2006 ImageDetail 50km by 60km

Classification probability (forest)Dark green High probabilityLight green ‘uncertain’


Classification probabilities (forest)Dark green High probabilityLight green ‘uncertain’No classification in cloudy area… Lead to use radar data instead

Digital Classification of Cloudy Image

NCAS 2005 Image + cloud mask (blue lines)Detail 50km by 60km


Eigenvector-Based Polarimetric Target Decomposition

Eigenvectors / Eigenvalues Analysis

T U U u u u u u u

T

3 3

11 2 3

1

2

3

1 2 3

0 0

0 0

0 0

*

OrthogonalEigenvectors

Real Eigenvalues

Pii

kk

1

3

Polarimetric Entropy

Probabilities

1

logn

i n ii

H P P

cos sin cos sin sini iTj j

i i i i i iu e e

3

1i i

i

P

Unit Target Vector

alpha

(Cloude-Pottier, TGARS, 1997)


Entropy/alpha (H / Space

LowEntropy

MediumEntropy

HighEntropy

SurfaceScattering

VolumeScattering

MultipleScattering

0 0.2 0.4 0.6 0.8 10

10

20

30

40

50

60

70

80

90

Entropy (H)

Alp

ha (a

)

93 X-BRAGG SURFACE

DIPOLE

DIHEDRAL SCATTERER

SURFACE ROUGHNESSPROPAGATION EFFECTS

FORESTRY DBLE BOUNCE

BRANCH / CROWNSTRUCTURE

CLOUD OF ANISOTROPICNEEDLES

NON FEASIBLEREGION

1 4 7

2

6

5 VEGETATION8


Dual Polarisation Radar Mode

For reasons of cost, data rate and coverage in radar design, it often employs a single transmitted polarisation state and a coherent dual channel receiver to measure orthogonal components of scattered signal.

The PALSAR sensor is just such a fully coherent-on-receive mode.

Such dual polarised radars are not capable of reconstructing the complete scattering matrix [S] but instead can be used to reconstruct a 2x2 wave coherency matrix [J].

(Cloude, POLINSAR 2007)


Dual Polarised Entropy/alpha Decomposition

**

**

][HVHVHHHV

HVHHHHHHH

SSSS

SSSSJ

**

**

][VHVHVVVH

VHVVVVVVV

SSSS

SSSSJ

Wave Coherency Matrix

(H Transmit, H,V Coherent Receive - PALSAR)

(V Transmit, H,V Coherent Receive -)




Related to 3x3 Polarimetric Coherency Matrix [T]

33*

2313

2313122211* Re2

2

1

10

01

01

100

011

2

1

100

011

2

1

ttt

tttttTkkJ

kk

THHH

H

33*

2313

2313122211* Re2

2

1

10

01

01

100

011

2

1

100

011

2

1

ttt

tttttTkkJ

kk

TVVV

V




cossin

sincos

0

0*

2

2

121

][i

i

e

eU

P

PDyyxy

xyxx

JJ

JJJ

2

122

221212

log

22

iii PPH

PPPPP

Scattering Angle and Entropy



Dual Polarised Entropy/alpha Space

Entropy H

alp

ha (

degre

es)

... Genuine decomposation classess to be inverstigated



Alpha (left) and Entropy (right) Maps of PALSAR Scene 381-6340 Acquired on 4 Oct 2008


AOI: a 10x10km Square (yellow box)

alpha/Entropy/Intensity Forest/Non-forest


Partial Polarimetric Coherence Optimisation

The coherence between two different polarisation channels:

jjj

T

jiii

T

i

jij

T

i

jiij

TT

**

*

,~

ji

where <> denotes spatial averaging, and contain the polarimetric information, while contain baseline dependent polarimetric and interferometric information.

iiT jjT jiij

In the HH-HV pair, where and , total decorrelation over the forested areas is observed since the predominantly polarimetric decorrelation between the HH-HV polarised backscattered signals is from areas dominated by volume scattering.

Tj 0,2/1,2/1 Ti 0,2/1,2/1

(Cloude & Papathanassiou, 1997)

According to Reigber et al. (IGARSS 2008), HH-HV is clearly the better choice for all forested areas.


Polarimetric Coherence OptimisationTo solve the coherence optimisation problem, we must maximise the modulus of a complex Lagrangian function L defined as

11 *** jjj

T

jjiii

T

iijij

T

i TTL

0

0

**

*

*

*

jjjjiTijT

j

T

iiiijijT

i

TL

TL

ijiiTijjjijii

jjijijiiTijjj

TT

TT

**11

*1*1

The maximisation problem can be described by setting the partial derivatives to zero.

By solving these matrix equations, the estimates for and the optimal scattering mechanisms and the corresponding coherences in images i and j are obtained from the resulting eigenvalue problems

i j

(Cloude & Papathanassiou, 1997)


Partial Polarimetric Coherence Optimisation

0.9

0.1

0.5

HV Coherence

Optimised Coherence


Joint Processing of Dual Polarised Entropy/alpha Decomposition and Partial Polarimetric Coherence

Optimisation 1). Generation of Entropy/alpha maps from PALSAR FBD SLC data implementing the above dual polarised Entropy/alpha decomposition algorithms;

2). Creation of the forest/non-forest discrimination map/mask using the Entropy/alpha classifier;

3). Coherence optimisation using multiple scattering mechanism approach described;

4). Non-forest region removal from the coherence map by the forest/non-forest mask derived from dual-pol Entropy/alpha maps;

5). Verification by in situ LiDAR forest canopy height data.


Optimised Coherence and LiDAR Canopy Height Map

Optimised Coherence of AOI: a 10x10km Square with Non-forest Mask in White of HH-HV Pair Acquired on 19 Aug and 4 Oct 2008

LiDAR Forest Canopy Height Map of the AOI Acquired in Sept 2007: Blue – 0 meter, Green- 20 metres and Red – 40 metres. (Courtesy of Forestry Tasmania)


Optimised Coherence and LiDAR Canopy Height Map

According to Le Toan (K&C 2010), the interferometric coherence ratio is sensitive to forest canopy height and the trend of coherence ratio is decreasing with a canopy height increase. That means the low coherence (red in left image) indicates a higher canopy height (red in right image). One reason for anomaly/inconsistency in some areas could be that the ground cover changed due to the different acquisition dates of radar and LiDAR data (one year gap).


Conclusions and Future Work

Based on dual polarised Entropy/alpha decomposition and the partial polarimetric coherence optimisation, an integrated forest/non-forest discrimination and optimised coherence–forest height estimation method for producing forest extent change and trend information was proposed.

Initial results on (a) the use of dual polarised Entropy/alpha decomposition for forest/non-forest discrimination, (b) optimised coherence of PALSAR dual-pol data as a source of information for forest height retrieval are consistent with in situ LiDAR forest height data.

Future work will aim at quantitative analysis of accuracy of forest/non-forest discrimination and relation of coherence-forest height with help of reference data.

Contact UsPhone: 1300 363 400 or +61 3 9545 2176

Email: [email protected] Web: www.csiro.au

Thank you

Zheng-Shu ZhouCSIRO Mathematics, Informatics and Statistics

Phone: 08 9333 6189Email: [email protected]: www.cmis.csiro.au

pres_h2a_okconvert.ppt

Technology

dual polarised radars

coherence optimisation

entropyalpha classifier

coherence map

polarimetric information

dual polarisation radar

entropy h alpha

coherent dual channel