overview of research activities at australia's national ... · measuring soil moisture 1)...

Overview of Research Activities at Australia's Overview of Research Activities at Australia's

National Science Organisation (CSIRO)

GENEVA SEMINAR – MAY 14, 2013

DIVISION OF MATHEMATICS, INFORMATICS & STATISTICS

Dr. Eric A. Lehmann | Research Scientist

Presentation Overview

• CSIRO as a national research organisation• Some quick facts and figures

• Combined radar–optical data for forest mapping and monitoring• Research background: international initiatives, optical and radar datasets

• Canonical Variate Analysis & Maximum Likelihood Classification

• Joint processing with Bayesian Conditional Probability Network

Research Activities at Australia's CSIRO | Eric A. Lehmann2 |

• Joint processing with Bayesian Conditional Probability Network

• Model–data fusion for water resources assessment• Background: WIRADA project, soil moisture data

• Data assimilation, data blending and evaluation

• Bayesian hierarchical modelling

• Fine-scale monitoring of complex environments using aerial and other spatial data (if time allows)

Australia’s CSIRO – Facts and FiguresCommonwealth Scientific and Industrial Research Organisation• Australia’s national research agency, one of the largest in the world

• provides scientific solutions to industry, governments and communities in Australia and worldwide

• established in 1926, now ~6’000 employees, 55 sites throughout Australia and overseas, including:

• Australia Telescope at Parkes, NSW

• research vessel Southern Surveyor

• laboratory in France


• laboratory in France

• field station in Mexico

• Multi-disciplinary research activities:

• agribusiness

• energy and transport

• environment and natural resources

• health

• information technology

• telecommunications

• manufacturing, mineral resourcess, etc.

Australia’s CSIRO – Facts and FiguresOutcomes-driven research:


Advantages for researchers: involved in many different application areas, e.g.

• seabed condition mapping using underwater acoustic echo sounding data

• forest and sparse vegetation mapping (National Carbon Accounting System)

• urban landscape monitoring with aerial photography

• combined radar–optical data for forest monitoring

• data assimilation for water resources assessment and accounting

• modelling of extreme weather events

• etc.

Combined analysis of optical and radar remote sensing data for forest Combined analysis of optical and radar remote sensing data for forest mapping and monitoring

CSIRO Mathematics, Informatics & Statistics, Perth, Australia

Cooperative Research Centre for Spatial Information (CRC-SI), Sydney , Australia

Landcare Research, Lincoln, New Zealand

5 | Research Activities at Australia's CSIRO | Eric A. Lehmann

BackgroundAssess and take advantage of the complemen-

tarity of synthetic aperture radar (SAR) and

optical sensors for forest/non-forest (F/NF)

mapping and monitoring

Motivation

• technological advances in synthetic aperture radar (not cloud-affected)


• technological advances in synthetic aperture radar (not cloud-affected) complement the existing optical datasets

• GEO-FCT: Forest Carbon Tracking task of the Group on Earth Observations (in support of global forest carbon estimation)

• Australia’s response to GEO-FCT: International Forest Carbon Initiative (IFCI) to increase forest monitoring capacity

• further development of the National Carbon Accounting System (NCAS) developed by CSIRO & partners: continental Landsat-based forest monitoring system

Data and Study AreaPilot study area

• calibration site defined as part of Australia’s GEO-FCT demonstrator project under IFCI

• 3’300 km2 area in north-eastern Tasmania (currently processing whole of Tasmania – 68’400 km2)

• main land covers:


• main land covers:• dry & wet eucalypt forest

• non-eucalypt forest

• rainforest

• plantations / deforestation

• agriculture & urban areas

• significant topographic variation (elevation: 80m to 1500m)

Datasets for F/NF mapping

PALSAR (HH,HV,HH-HV)

ALOS-PALSAR• fine-beam dual polarisation (HH and HV),

L-band (23.6cm)

• ascending orbit (34.3° off-nadir)

• pre-processed to 25m pixel size

• acquired Sept./Oct. 2009

Landsat TM• 6 spectral bands (thermal band omitted),

25m pixel size

• from the NCAS archive of MSS/TM/ETM+ imagery

• acquired Jan. 2009

Landsat TM (bands 5,4,2)


Radar–Optical F/NF Classification

Step 1: define spectral classes using Canonical Variate Analysis (CVA)

268 training sites selected for the classification,

representing a broad range of landcover types

over the study area

Analyses carried out for:

1. Landsat data (6 bands)

2. PALSAR data (2 bands)

05

1015

CV

2

169

170172173

174

180

181

182183

184185

188

189190191193195

196197198

199

200201202

203204

205

206

207

208

209

210

211

212 213

214

217

220

221222 223

225

228

230231 232

234

235

236

237

239

240241242

243

244

245

246

247

248

249251

253

254

255

256

257

258

259

262

265

266267

2681

2

3

4

5

6

7

89

10

11

1213

14

15

16

171819

20

2122

23

24

25 26

27

282930313233

3435 3637

38394041

42

43

44 454647

48

49

50

5152

53

54

55

565758

5960

6162

636465

66

67

6869

70

7172

73

7475

7677

78

79

80

81

83

84

85 86

87

88

89

90

9192

93

9495

9697

98

99

100

101102103

104105

106

107

108

109

110

111112

113114115

116

117

118119

120

121

122123124

125

126

127

128

129130

131

132

133

135136

137

138141

143

144

145

146 147

148

149

150

151

152153

154

155

156

157

158

159

160

162164

166

agriculture (crops)

forest (dense)


2. PALSAR data (2 bands)

3. combined SAR–optical data

(8 bands, concatenation)

Canonical roots (measure of separability):

1. Landsat only:18.9 8.6 3.2 1.7 1.1 0.5

2. PALSAR only: 24.0 2.9

3. combined PALSAR–optical data:28.9 12.6 7.7 3.3 1.8 1.3 1.1 0.4

Training sites in CV1-CV2 space for

Landsat data (4 out of 7 classes

shown). Colour legend: forest sites,

non-forest sites, cleared/immature

plantations.

0 5 10 15 20 25 30 35

-50

CV1

171

175176

177178179

186187

189190192

194

206

215

216

217

218 219224

226

227

229

231 232

233

235

238

250252260

261263264

82

100 121134 139

140

142161

163

165

167168

water

bare ground

Definition of spectral classes using CVA

The number of selected sub-classes reflects the ability of each

dataset to discriminate between different land covers



Step 2: Maximum-Likelihood Classification (MLC), using the spectral classes defined by CVA

Example: Ben Lomond region, alpine heathland (shrubs)


PALSAR (HH/HV/HH-HV) Landsat (bands 5/4/2) TASVEG reference


SAR F/NF classification Landsat F/NF classification SAR–Landsat classification

5km

Multi-Temporal Radar–Optical Processing• Assume that the datasets are not coincident temporally

• Consider independent forest probability maps from each dataset

• Refinement of the single-date forest classifications using a Bayesian Conditional

Probability Network (CPN): spatial-temporal model, hidden Markov model

Landsat

prob. image 1972LandsatLandsat

Landsat

prob. imageLandsat

SAR prob.

image 2006

Landsat time series (NCAS) SAR–Landsat time series

1972


2012

LandsatLandsat

prob. image

CPN

forest map1972

2012

forest map

2012

Landsatimage 2006

CPN

forest map1972

2012

forest map

Landsat

prob. image

Landsat

prob. image

Multi-Temporal Radar–Optical ProcessingCombined multi-temporal forest map result for 2006 (CPN outputs):

green layer: 2006 Landsat-only (binary) forest map

red layer: 2006 SAR-Landsat (binary) forest map → ~95% idenKcal


Combined Radar-Optical Processing

Summary

• SAR and optical sensors provide complementary information for forest mapping and monitoring

• quantify improvement of F/NF classification achieved by jointly considering SAR and Landsat:

◦ adding SAR bands to the optical data provides one additional dimension for


◦ adding SAR bands to the optical data provides one additional dimension for classification

◦ L-band SAR data allows more separation than C-band

◦ with SAR, the cross-polarisation (HV or VH) provides most of the discrimination information

• strategies for dealing with non-coincident datasets:

◦ use of a multi-temporal approach (e.g. conditional probability network)

Bayesian hierarchical modelling for water resources assessment and accounting

Eric Lehmann & Grace Chiu



National water accounting and assessment under WIRADA / AWRA

• Water Information Research and Development Alliance

• Alliance between CSIRO and the Bureau of Meteorology (BoM)

• Monitor status of Australia’s water resources + forecasting of availability

• Australian Water Resources Assessment (AWRA): BoM activity component, system of models, model-data fusion

• AWRA-L: landscape hydrological model for AWRA

Background


• AWRA-L: landscape hydrological model for AWRA

• Observational data in AWRA-L: used in model development, (global) parameter estimation, forcing (e.g. precipitation)

• Additional datasets exist with new/other characteristics

⇒⇒⇒⇒ Need to reconcile or integrate observed and modelled estimates.

Research Focus

•Soil moisture (SM)

• one of the possible variables of interest

in WIRADA / AWRA

• availability of SM products:

• ground-based

• remote sensing

• case-study: could be replaced with any


• Murrumbidgee River Catchment (MRC)

• 73’400 km2, southern NSW, Australia

• availability of ground probes for “benchmark” SM measurements (OzNet monitoring network)

→ case-study with aim to up-scale nationally

• case-study: could be replaced with any

other variable...

Measuring Soil Moisture

1) In-situ ground probes: OzNet, “point-level” SM at depth of ~0–5cm.

2) Remote sensors, e.g. AMSR-E, ASCAT, ASAR, etc.: deterministic retrievals from brightness temperature, SM at depth of about 1–2cm.


about 1–2cm.

3) Physical models: e.g. AWRA-L, CABLE, etc. → ... not considered in our preliminary model.

⇒⇒⇒⇒ Different temporal & spatial resolutions!

Data Assimilation⇒ How to reconcile/consolidate SM products and assess uncertainty?

Data assimilation: via Kalman filter, particle filter, 3D-KF, etc.

• model-based temporal smoothing

• typically ignore spatial correlation (no spatial smoothing), or use non-model-based estimation of spatial correlation

• usually require “manual” alignment of pixels (space) and time intervals


Bayesian Hierarchical Modelling

⇒⇒⇒⇒ How to reconcile/consolidate SM products and assess uncertainty?

Probabilistic inference: estimate of posterior density (Bayes’ rule!)

p(X|y) ∝ p(y|X) ⋅ p(X)

↑posterior ↑likelihood ↑prior

where: X is the latent state variable (and parameters)


where: X is the latent state variable (and parameters)

y is the data (instantiation of random variable Y)

Estimation of posterior density leads to estimates of:

• posterior mean / median (or mode)

• credible intervals (Bayesian equivalent to confidence interval) → uncertainty estimates!


⇒⇒⇒⇒ How to reconcile/consolidate SM products and assess uncertainty?

Probabilistic inference: estimate of posterior density (Bayes’ rule!)

p(X|y) ∝ p(y|X) ⋅ p(X)

↑posterior ↑likelihood ↑prior

Proposed approach: statistical spatio-temporal modelling


Proposed approach: statistical spatio-temporal modelling

• model-based temporal smoothing

• model-based spatial smoothing

• model-based spatial alignment

• model-based imputation for missing data

• single hierarchical model for unified inference

... “temporal” aspect not considered in current (preliminary) model!

Proposed Bayesian Hierarchical Model

Model-based spatial smoothing: at fixed time t

Product q

Product p

ground

probes

AMSR-E


Covariate x

(driver of SM)

AWAP

(precip.)

08/01/2007

time

Model-based spatial alignment (preliminary model): linking the spatial datasets at different resolutions


Product qProduct p

(remote sensing)RESPONSE


⇒⇒⇒⇒ Aim: benchmark AMSR-E product vs. probes...

(precipitation) covariate x

State s

DRIVER

Preliminary spatial model for SM (Murrumbidgee Catchment): every quantity is related to each other through a single model

ground probes:

AMSR-E SM:

latent SM:

AWAP:



spatial patterns:

explicit modelling of

spatial correlation


Conditional auto-regression (CAR): a form of 5th nearest-neighbour dependence with exponential decay

→ models spaKal dependence

beyond one but less than two

AMSR-E pixels (so that SM

state is representative of


state is representative of

AMSR-E pixels)

AMSR-E pixel

AWAP pixel


Given the model structure, the latent soil moisture s (and other variables) is estimated / fitted via MCMC sampling (Metropolis-Hastings within Gibbs):

• ~105 to 106 iterations, basic convergence diagnostics

• super-computing facilities at CSIRO (CPU & GPU clusters)

• parallelised implementation in R, some computationally intensive routines coded in C/C++


intensive routines coded in C/C++

• currently ~1M iterations per 24h.

• Recall model hierarchy (at fixed time):

(p,q) ← s ← x ← φ

• Model fit for OzNet ground probes: and q

Spatial Model Fit: 18/01/2007


p,


Model hierarchy (at fixed time): (p,q) ← s ← x ← φ• Estimate of latent SM : visible

influence of AMSR-E, AWAP & OzNet

• Inferring missing AMSR-E pixels : some residual bias apparent (due to influence of precipitation)

• Posterior mean of spatial random effects : spatial autocorrelation


p,

x

effects : spatial autocorrelation clearly visible (motivation for proposed framework!)

• Latent soil moisture :


• 95% credible interval (CI): model-based estimate of uncertainty!(for 15 pixels in above map)

... issue with SM<0 in C.I. due to assumptions of

instruments specs (e.g. probe accuracy of ± X units)


• Model-based comparison of AMSR-E “performance” vs. benchmark:

• Interpret as: “AMSR-E is less precise than m-th probe by factor Rm”

Evaluating AMSR-E vs. Ground Probes


AMSR-E is

more precise...

08/01/2007

20/01/2007


Summary

Preliminary hierarchical model for soil moisture over the Murrumbidgee River Catchment:

• demonstration of statistical modelling framework (work in progress)

• unified model-based inference for SM assimilation & evaluation


Future developments:

• addition of temporal component

• look at further / other datasets (OzNet contribution to SM map is minimal)

• extension to larger and/or national scale

• faster code implementation

• integrate modelled estimates (e.g., AWRA-L) → model–data fusion

Urban Monitor: fine-scale monitoring Urban Monitor: fine-scale monitoring of complex environments using aerial and other spatial data

Mathematics for Mapping and Monitoring Group



Urban Monitor ProjectFine-scale monitoring of complex environments using remotely-sensed aerial and other spatial data

Pilot area over Greater Perth, WA:

• 9600 km2, yearly acquisitions since 2007

• PAN ∼0.1m GSD, multispectral ∼0.3m GSD

• 60% fwd overlap, 30% side overlap• 60% fwd overlap, 30% side overlap

• 35,000 frames , 13 – 40 TB per year

Issues for monitoring:

• geometric and radiometric calibration

• data processing and analysis

• storage


Urban Monitor ProjectFine-scale monitoring of complex environments using remotely-sensed aerial and other spatial data

Monitoring opportunities:

• wetland condition

• weed invasion, disease spread

• vegetation condition, tree density & growth• vegetation condition, tree density & growth

• river foreshore condition

• land use changes

• hydrological modelling, irrigation areas

• digital terrain model

• unauthorised clearing & water use

• etc.


Urban Monitor ProjectRadiometric calibration using calibration targets deployed each year


Urban Monitor ProjectRadiometric calibration through empirical statistical models, using BRDF kernels, gain + offset coefficients, and ground targets


Urban Monitor ProjectRadiometric calibration through empirical statistical models

Raw data Calibrated


Urban Monitor ProjectnDEM processing:

DSM

Ground

candidates


candidates

DEM

nDEM

Urban Monitor ProjectGenerating information for monitoring using:

• time-series DEM/DSM data + multi-spectral calibrated imagery

• two-stage CVAR (Canonical Variate Analysis with Rational polynomials) – supervised, hierarchical


Urban Monitor ProjectSpectral classification + nDEM leads to base classification: monitoring requirements vary according to local issues and stakeholders’ needs

• grass• trees• brown dirt• brown roofs• black tar• concrete ground• concrete roof• grey ground• grey ground• grey roof• shadow• pools


Thank youCSIRO Mathematics, Informatics & StatisticsEric A. Lehmann

t +61 8 9333 6123e [email protected] www.csiro.au

EARTH OBSERVATION INFORMATICS TCP

Thank you

overview of research activities at australia's national ... · measuring soil moisture 1)...

Documents