close

Anmelden

Neues Passwort anfordern?

Anmeldung mit OpenID

GEO 214 Vorlesung Wie nutze ich Radardaten?

EinbettenHerunterladen
GEO 214
Vorlesung
Wie nutze ich Radardaten?
•
•
•
•
Speckle
Klassifikation von Radar Bildern
Informationsgewinn aus Radarbildern (Modellierung)
Anwendungsbeispiele
Radio Detection And Ranging RADAR
•
Laufzeit:
•
Entfernung:
t = 2R
c
R = ct
AKTIVES System mit
eigener Energiequelle
Unabhängig von
Tageszeit
2
1. Amplitude
2. Phase [0, 2π]
0 bzw. 2π
0 bzw. 2π
0,5π
Disputation Thiel
Sensor
Sensor
Measured phase is dependent on
Distance
Streuobjekt
Streuobjekt
Disputation Thiel
Sensor
Measured Phase is dependent on
Object Characteristics
Streuobjekt
Streuobjekt
Streuobjekt
Disputation Thiel
Phase erscheint zufällig
Das ist sie aber nicht!!!
Speckle Filter
• Die Rückstreueigenschaften eines
Oberflächenelements können wegen
Speckle nicht durch ein Pixel ermittelt
werden
• Rückstreukoeffizient muss über mehrere
Pixel geschätzt werden
Auflösungsverlust
• Statistische Eigenschaften des Speckle
und der Rückstreuung selbst müssen
beachtet werden:
Gamma MAP filter:
Frost filter
(adaptiver Mittelwertfilter)
Problem: Textur in Schätzfenster
Multichannel Filtering
Linear Combination of k channels increases ENL with reduced cost of
resolution (Quegan & Yu, 2001):
−1
C σ
Aki = I k t I −1
σ CI σ
M
J k = ∑ Aki I i
i =1
A=weighting coefficient,C=covariance matrix,
σ=mean vector (spatial estimate), J=filtered image
When ignoring ENL differences
between channels & correlation:
Jk =
σ k0
M
M
Ii
∑σ
i =1
0
i
Required spatial average from moving window or
previously Frost filtered images
Multichannel Filter
ENL = 3
ENL = 70
Test area A
60
80
Optimal Filter
Simple Filter
Theory
60
20
40
ENL
ENL
ENL
40
0
Test area C
Test area B
60
40
20
20
0
10
20
No. of Images
Acker
30
0
0
10
20
No. of Images
Wald
30
0
0
10
20
No. of Images
Acker
30
Problem: Korrelation des Speckle
Kohärenz für Bildpaare:
(a) 25 Dec. 93 / 31 Dec. 93,
(b) 25 Dec. 93 / 06 Jan. 94,
(c)
25 Dec. 93 / 27 Jan. 94,
(d) 25 Dec. 93 / 20 Feb. 94.
Wann arbeite ich in der
Logarithmischen-Skala (=dB)?
Radarrückstreuquerschnitt – in [dB]
Sigma Null [linear]
Sigma Null [dB]
ASAR HV – Sommer 2005
Radarrückstreuquerschnitt – in [dB]
[dB] = dezibel
Vorsicht –
unterschiedliche Statistik
• Wie schätze ich den
Rückstreuquerschnitt?
Nur im linearen Bild –
Mittelwert aus mehreern
Pixeln eines homogenen
Objekts ist bei GammaVerteilung der beste Schätzer
(nicht der Fall für Verteilung
der dB Daten).
• Wann nehme ich dB Skala?
Visuelle Interpretation
Maximum-Likelihood-Klassifikation, da
Varianz gegenüber linearen Daten
stabiler
Es ist wichtig, die Bildstatistik zu
kennen
Annual Mean σ 0
MVA [dB]
150
800
100
Cropland
ENL=5
Forest
600
ENL = 5
400
Grassland
ENL=9.1
ENL = 8
2
4
6
MVA [dB] VV pol.
0.05
8 σ
0
0
600
400
200
0
MVA [dB]
800
Forest
50
0
Annual Mean σ 0
MVA [dB]
ENL = 70
200
0.1
0
2
4
6
MVA [dB] VV pol.
8
0.05
0
σ
0
0.1
2
4
6
MVA [dB] VV pol.
8
Radiometrische Anforderungen
Annual Minimum HV: Grassland vs. Cropland
ENLsolid =3 ENLdashed =50
σ0
μ = 20
σ1
log e μ
DP =
⎛1
1 ⎞
⎜⎜ − ⎟⎟
⎝ σ1 σ 2 ⎠
0
Error for each class:
δ class 2 = ∫
DP1
γ pdf (I , L,σ 10 / L )ΔI ⎞⎟
⎝
0
DP1
γ pdf (I , L, σ 20 / L )ΔI
0
50
0
2
0
0.01
0.02
σ
⎠
0
0
0.03
Rad. uncertainty [dB]
0
50
-2 100 ENL
Annual Minimum HV: Cropland vs. Forest
400 ENLsolid =3 ENLdashed =50
P(σ 0)
δ class1 = 1 − ⎛⎜ ∫
200
100
300
200
σ2 =-16.0055
0
100
0
100
0
σ1 =-20.3651
Error [%]
Decision Point:
0
σ1 =-20.9895
0
σ2 =-20.3651
Error [%]
P(σ 0)
400
50
0
2
0
0.01
0.02
σ
0
0.03
0
0.04
Rad. uncertainty [dB]
0
50
-2 100 ENL
Klassifikation von Radar-Daten
Maximum Likelihood Classification
Bayes Theorem:
Conditional
probability:
Post Classification
Iterated Contextual
Probability:
p{c}p{x | c}
p{c | x} =
p{x}
p{x | c} =
1
(2π )K / 2 Cc
1
2
⎛ 1
⎞
t
exp⎜ − (x − μ c ) Cc−1 (x − μ c )⎟
⎝ 2
⎠
⎛ ∑ pδ (c x ) ⎞
⎜
⎟
pn (c ) = ⎜ δ
⎟
n
δ
⎜
⎟
⎝
⎠
β
Courtesy of Dorothea Kolossa – TU
Berlin
ICP
Classification Tree
Pruned classification tree
Vorteil: keine Normalverteilung vorausgesetzt
Land Cover Mapping by means of C-Band SAR
ASAR HV May 11th 2005
Fre q u e ncy
1500
Water
Grassland
Cropland
Forest
Settlement
1000
500
0
-25
-20
-15
Intensity [dB]
-10
-5
Even basic classes can hardly be distinguished having only one image and polarization
• Teilweise nur
noch
Topographie
erkennbar
JERS
Wald/NichtWald erkennbar
ERS-Szene
10 Juli ERS-2 VV
07 Dec. ERS-2 VV
Multitemporale Datenanalyse für Cband
Mean annual variability (Ii,j=intensity for
date i and j) (Quegan et al., 2000):
N −1
⎡
⎤
2
mva = 10 ⋅ log ⎢
R
∑∑
ji ⎥
(
)
1
−
N
N
i =1 j > i
⎣
⎦
I ⎞
⎛I
Rij = max⎜ i , j ⎟
Ii ⎠
⎝ Ij
Multi-temporal metrics of 14
ASAR AP HV pol. intensity
images with (a) MVA, (c)
annual minimum σ0, (d)
annual mean σ0. (b) shows
a Landsat ETM+ image
acquired Sep. 4th 1999
(NIR, Red and Blue
channel).
VV - Annual Max. VV - Annual MVA VV - Annual Mean
VV - Annual Min
Frequency
GL
AL
FO
ST
WT
All
-10 -5
0
5
0
5
10
-10 -5
0
-10
HH - Annual Min
Frequency
HH - Annual Max. HH - Annual MVA HH - Annual Mean
-20
-10 -5
0
5
0
5
10
-15 -10 -5 0
HV - Annual Min
Frequency
HV - Annual Max. HV - Annual MVA HV - Annual Mean
-25-20 -15 -10 -5
-20 -15 -10
0
σ [dB]
0
5
0
σ [dB]
-20
-10
0
σ [dB]
-25 -20 -15
0
σ [dB]
Textur
• Textur auch in Radar mit besonderen
Eigenschaften
Genauigkeit und ENL Anforderungen an
Klassification
50
0
0
50
ENL
50
0
100
0
50
ENL
100
Prod. Accuracy classes[%]
Accuracy [%]
κ
100
100
Prod. Accuracy classes[%]
HV pol.
100
100
50
0
0
50
ENL
100
0
50
ENL
100
VV pol.
100
Accuracy [%]
κ
100
50
0
0
50
ENL
100
50
0
0
50
ENL
100
50
0
Modellierung
LE TOAN et al. 2001: 4
•
NASA JPL AIRSAR
airborne results
using polarimetric
C-, L- and P-Band
data with incidence
angles between
40° and 50°
degrees.
IMHOFF 1995: 514
Volume scattering
NASA SIR-C Science Plan, 1988
Angular variation of stalk
absorption loss factor for
wheat at 5 GHz (Ulaby
Et al. 1986: 1566).
Angular response of scattering
coefficients for 3 wavelengths
5 fields with high levels of moisture content
L-Band
C-Band
X-Band
[Ulaby et al. 1986:1825]
Rückstreuung von Getreide im Laufe der Pflanzenentwicklung
HV
HH
0
-4
Rape
Winter cereals
Maize
Potatoes
Summer cereals
Sugar beet
-6
-8
-10
-5
0
σ [dB]
0
σ [dB]
-12
-14
-16
-10
-18
-20
Rape
Winter cereals
Maize
Potatoes
Summer cereals
Sugar beet
-22
-24
-15
22/04 12/05 01/06 21/06 11/07 31/07 20/08 09/09 29/09 19/10 08/11 28/11
Date
22/04 12/05 01/06 21/06 11/07 31/07 20/08 09/09 29/09 19/10 08/11 28/11
Date
VV
0
150
Maize
Rape
-2
Wheat winter
Sugar beet
Wheat summer
Potatoe
-4
Height [cm]
0
σ [dB]
100
-6
-8
50
-10
-12
Rape
Winter cereals
Maize
Potatoes
Summer cereals
Sugar beet
01/05 21/05 10/06 30/06 20/07 09/08 29/08 18/09 08/10 28/10 17/11 07/12
Date
0
12/04
22/04
02/05
12/05
22/05
Date
01/06
11/06
21/06
ESA Specialist Panel on Agriculture, SP-1185, October 1995
Multitemporal ERS-1-Composite, UK. R: 25. May, G: 13. June, B: 29. Juni.
W: Winter Wheat, R: Rape Seed, WB: Winter Barley, G: Grassland
Source: Zmuda et al. 1994
Time:
Time Series ASAR AP mode, C-Band HV backscatter.
Test area: Bolshe-Murtinsky, Siberia.
M. Santoro 2004
AP - HH,
Polarizations:
AP - HV,
IS1
IS1
01.03.03,
01.03.03,
50x50m,
50x50m,
IM, VV
IS2
WS, VV
27.02.04,
21.11.03,
50x50m
75x75m
Incidence angle:
IS2
IS7
ASAR IM,
Siberia
Backscatter and coherence
A
B
C
D
Layer 2
Layer 1
At C-band (5 cm) A and B are the main scattering mechanisms
At L-band (25 cm) the wave penetrates more into the canopy
Coherence is determined by the temporal stability of the forest
M. Santoro 2004
Austrian pine
X band
λ= 3 cm
L band
λ= 27 cm
P band
λ= 70 cm
VHF
λ>3m
Dichter Wald:
Wald
Sensing „vegetation“
Offene Flächen:
Sensing „ground“
M. Santoro 2004
Forest inventory data
Accurate forest inventory data is a necessity
Problems: Better data for forests of high economic value. Bad quality in
remote areas.
Scatterometerprofil einer Kiefer:
σ 0 for = σ 0 gr + σ 0 db + σ 0veg
σ
h
0
veg
= ∫ σ v ⋅ e −α ⋅(h − z )dz
0
o
(1 − Ttree )]
σ ofor = (1 − η )σ gro + η [σ gro Ttree + σ veg
Water Cloud Model
A
B
C
D
Rückstreuung über Wald
σ 0 for = σ 0 gr + σ 0 db + σ 0veg
Rückstreuung von
Baumkrone
σ
h
0
veg
= ∫ σ v ⋅ e −α ⋅(h − z )dz
0
Basierend auf „radiative transfer“:
Energietransfer durch eine
Wasserwolke.
Water Cloud Model incl. gaps
σ
o
for
= (1 − η )σ
o
gr
[
+η σ T
o
gr tree
+σ
o
veg
(1 − Ttree )]
A water cloud with gaps is closer to reality and easy to handle
o
(1− e−βV )
σ ofor = σ gro e−βV + σ veg
Vegetation backscatter
Ground backscatter
Forest transmissivity
Vegetation backscatter
Forest transmissivity is related to area-fill and tree attenuation
Water Cloud Model
1.
Model training / regression
Backscatter only: Estimate σ 0veg, σ 0gr and β using (σ ο, V) measures.
ERS
o
(1− e−βV )
σ ofor = σ gro e−βV + σ veg
JERS
Retrieval: JERS
backscatter
RMSE: 33 m3/ha,
Relative RMSE: 22 %
Santoro et al., RSE, 2002
P2
B
P1
© CCRS - Fundamentals of Remote Sensing Tutorial
Interferometry
α
θ
Bp
.
Bs
Phase difference because of running
time difference
Bp
d2
h0
d1
T
hT
Phases of the single measurements
Phase C-VV
Phase C-VV
Bamler, R., InSAR
Sommerschule 2002
Across-Track SAR
Interferometry
•
phase of complex pixel in …
– … SAR image #1:
4 ⋅π
φ1 = −
⋅ R + φ scatt ,1
λ
– … SAR image #2:
φ2 = −
4 ⋅π
λ
⋅ (R + ΔR ) + φscatt , 2
– … interferogram:
φ = φ1 − φ2 =
•
4 ⋅π
λ
⋅ ΔR
phase-to-height sensitivity:
(if φscatt ,1 = φscatt , 2 !)
Coherence !!! Speckle
∂φ 4 ⋅ π
B⊥
=
⋅
∂z
λ R sin Θ
[R. Bamler, MFFU Sommerschule 2000]
Coherence and InSAR phase
γ=
s2 s1 *
s1s1 * s2 s2 *
complex interferogram
γ = γ e − iφ
γ =
φ=
degree of coherence
interferometric phase
= ensemble average
s2 , s1 = co-registerd complex image values
[Strozzi, InSAR Sommerschule 2002]
Phase noise - examples
γ = 0,28
γ = 0,5
γ = 0,65
γ = 0,82
Coherence ASAR-IMS HH
April/Mai
April/August
Thüringen
Thüringen
ERS-1/2 tandem coherence
Weather effects
Strong
decorrelation
occurs with
rainfall
Temporal Decorrelation
ERS tandem
(1 day)
ERS long-term
(35 days)
[Strozzi, InSAR
Sommerschule 2002]
Causes for decorrelation
• Temporal decorrelation
•
•
•
•
•
Seasonal variations
Temperature
Precipitation
Snow cover
Wind speed
• Volume decorrelation
Environmental conditions
Change detection
JERS Coherence 1993-12-29 – 1994-02-11
JERS Coherence 1996-01-17 – 1996-03-01
Leif Eriksson, ForestSat 2005
Change detection – Method 1
- Requires stand borders
- If the circle appears on the
line no change occurred
between December 1993 and
February 1996
- Clear-cuts show
increased coherence
- Burned forest still
under evaluation
Leif Eriksson, ForestSat 2005
Vegetationsstruktur
Species
Height
Stem Diameter
Biomass
(Stem Volume)
Age
Structure
Chunsky and Bolshe
Stem Density [/ha] log
☺
Stem Density [/ha]
5000
4000
3000
2000
1000
0
0
20
40
DBH [cm]
60
15
r=-0.993
r=-0.989
10
5
0
0
ln(N)=-1.621*ln(DBH)+11.65
1
2
3
4
DBH [cm] log
5
Forest structure / Quality of inventory data
r = -0.746
RS > 50 %
r = -0.895
Area > 3 ha
r = -0.678
RS > 30 %
r = -0.746
Area > 3 ha
Failed update
of inventory
data
(Santoro et al. 2007)
Coherence modelling
1
2
3
Layer 2
Layer 1
Interferometric Water Cloud Model (IWCM)
0
σ gr0 − βV
σ veg
(
α
e − jωh − e −αh )⎤
− βV ⎡
γ for (V ) = γ gr 0 e + γ veg 0 (1 − e )⎢
⎥
−αh
(
)
(
)
σ for
σ for
α
j
ω
1
e
−
−
⎣
⎦
Ground contr. (1+2)
Vegetation contr. (3)
Forest complex coherence, γfor
Γveg, Vegetation contribution
ϕfor
ϕveg
Γgr, Ground contribution
2.
Testing
Invert the model using backscatter and/or coherence values to
estimate stem volume.
Error
1. ERS backscatter does not provide any information
60 %
2. JERS backscatter provides rather good results
45 %
☺
Dry-unfrozen conditions
Winter-frozen conditions / weather changes
3. ERS „tandem“ coherence provides best results
30 %
Local survey
15 %
Accuracy depends on
1) weather conditions, 2) inventory unit
Santoro et al., RSE, 2002
3.
Multi-temporal combination
ERS „tandem“ coherence
RMSE: 10 m3/ha
Relative RMSE: 7 %
JERS backscatter
RMSE: 33 m3/ha,
Relative RMSE: 22 %
Santoro et al., RSE, 2002
Beispiel für großflächige
Anwendung
Mat. For
est
Med. Fo
rest
Clear Cu
t
ields
Smooth f
-13
-9
JERS sigma naught
-4
SAR Imaging for Boreal Ecology and Radar Interferometry Applications (SIBERIA)
-16
Water
Red: ERS coherence
Green: JERS Intensity
Blue: ERS-2 summer intensity
0.2
0.3
0.4
0.5
ERS coherence
0.6
0.7
Classification of stem volume using ERS tandem coherence and JERS
intensity data
ERS tandem data acquired during autumn 1997
JERS data acquired during summer 1998
Separability of classes
Histogram Analysis
Stem volume classes: 1
2
3
4
0-20 m^3/ha
20-50 m^3/ha
50-80 m^3/ha
>80 m^3/ha
Good correlation was found for γH, σH
and the mean value of the forest class
>80m^3/ha
Input for the Maximum Likelihood Classification
v
−
122.1
γ (v) = γ 75 + (0.330+ 0.581⋅ γ 75 ) ⋅ e
v
−
107.34
σ 0(v) =σ75 −2.46⋅ e
(WAGNER et al. 2003)
Independent Russian Field Data
Ground
validation
Earth
Obs.
Results
<20
<20
m3/ha
20-50
m3/ha
50-80
m3/ha
>80
m3/ha
Total
User
accuracy
908
36
5
9
977
93 %
20-50
m3/ha
76
576
39
15
707
81 %
50-80
m3/ha
12
33
881
58
984
90 %
>80
m3/ha
0
9
120
2182
2311
94 %
Total
1016
655
1045
2264
5232
95 %
Producer
accuracy
89 %
88 %
84 %
96 %
95 %
91%
m3/ha
Pooled confusion matrix for the 7 Russian test areas.
κw= 0.94.
SIBERIA Products
Radar Image Mosaic
111 Radar Image Maps
Forest Cover Mosaic
96 Forest Cover Maps
Document
Kategorie
Technik
Seitenansichten
2
Dateigröße
16 525 KB
Tags
1/--Seiten
melden