Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Prediction of productivity indices using remotely sensed data
Michael Watt, Jonathan Dash, Pete Watt, Santosh Bhandari
Introduction
• Site Index, 300 Index commonly used metrics to describe plantation productivity
• Typically estimates of these properties made at the stand level by foresters
• Estimates of these productivity surfaces have also been made using environmental variables
Introduction• Little research has investigated the utility of
remotely sensed data sources for predicting productivity metrics
• Use of remotely sensed data could : improve prediction accuracy allow fine scale spatial prediction
• Combinations of the following three data sources could be useful : Environmental Surfaces Satellite Imagery LiDAR
Data sources - Environmental surfaces
• Data mostly at 100 m2
resolution
• Air temperature• Solar radiation • Rainfall• Relative humidity• Slope• Soil fertility
MW2
Slide 4
MW2 Pete, Santosh, if you have any additional information or pictures please add these. Michael Watt, 4/03/2015
Data sources - satellite imagery
• RapidEye most useful satellite 5 m resolution, 0.01-0.02 $/ha
• 5 spectral bands• Vegetation ratios that describe
photosynthetic activity can be derived.
• Textural measures that describe vegetation surfaces can be derived.
• Easily incorporated into GIS
Data sources - Light detection and ranging (LiDAR)Active remote sensing technology
LiDAR emits ~ 200,000 pulses per secondRecords up to 4 returns per pulseEach return is converted to 3D point
Site Index from LiDAR
• LiDAR can be used to accurately predict height
• Using stand age and a age-height equation wecan estimate Site Index
• Site Index predicted from LiDAR was usedfor modelling
Objective
• Develop models of 300 Index and Site Index using different combinations of LiDAR, satellite imagery and environmental surfaces
• Models developed with and without stand age using both non-parametric and parametric methods
Methods
• LiDAR acquired early 2014 ~ 11 points m2
• RapidEye acquired January 2014
• Total of 493 plots installed early 2014
• 433 plots used for model fitting, 60 plots set aside for model validation
Methods
• Total of 14 models developed using various combinations of the three datasets.
• Parametric method used was multiple regression
• Non-parametric method used was k-nearest neighbour• Random forest or k-MSN used to assign neighbours• Values of k were optimised to minimise model error
Results - Site Index
Rap
idE
ye (R
E) s
pect
ral
RE
ratio
s
RE
text
ure
RE
spec
tal +
ratio
s +
text
ure
Envi
ronm
enta
l var
iabl
es
Env
ironm
enta
l var
iabl
es +
RE
LiD
AR
pre
dict
ed S
ite In
dex
(SI)
LiD
AR
Pre
d. S
I + E
nviro
n. +
RE
Roo
t mea
n sq
uare
err
or (m
)
1.0
1.5
2.0
2.5
3.0
ParametricNon-parametric k-NN
1.0
1.5
2.0
2.5
3.0
3.5
Models used to predict Site Index
Rap
idE
ye (R
E) s
pect
ral
RE
ratio
s
RE
text
ure
RE
spec
tal +
ratio
s +
text
ure
Envi
ronm
enta
l var
iabl
es
Env
ironm
enta
l var
iabl
es +
RE
LiD
AR
pre
dict
ed S
ite In
dex
(SI)
LiD
AR
Pre
d. S
I + E
nviro
n. +
RE
Coe
ffici
ent o
f det
erm
inat
ion
(R2 )
0.5
0.6
0.7
0.8
0.9
0.5
0.6
0.7
0.8
0.9
Without age
With age
Results - 300 Index
Rap
idE
ye (R
E) s
pect
ral
RE
ratio
s
RE
text
ure
RE
spe
ctal
+ ra
tios
+ te
xtur
e
Env
ironm
enta
l var
iabl
es
Envi
ronm
enta
l var
iabl
es +
RE
LiD
AR
pre
dict
ed S
ite In
dex
(SI)
LiD
AR P
red.
SI +
Env
iron.
+ R
E
Roo
t mea
n sq
uare
err
or (m
3 ha-1
yr-1
)
1.0
1.5
2.0
2.5
3.0
3.5
4.0
Parametric Non-parametric k-NN
1.0
1.5
2.0
2.5
3.0
3.5
4.0
Models used to predict 300 Index
Rap
idE
ye (R
E) s
pect
ral
RE
ratio
s
RE
text
ure
RE
spe
ctal
+ ra
tios
+ te
xtur
e
Env
ironm
enta
l var
iabl
es
Envi
ronm
enta
l var
iabl
es +
RE
LiD
AR
pre
dict
ed S
ite In
dex
(SI)
LiD
AR P
red.
SI +
Env
iron.
+ R
E
Coe
ffici
ent o
f det
erm
inat
ion
(R2 )
0.4
0.5
0.6
0.7
0.4
0.5
0.6
0.7
0.8
Without age
With age
Results – Best Models
Metric Age? Type Variables R2 RMSESite Index Y NP LiDAR + Environ. + RapidEye 0.91 1.40 m
N P Environmental + RapidEye 0.79 2.46 m
300 Index Y P LiDAR 0.79 2.45 m3 ha-1 yr-1
N NP Environmental + RapidEye 0.65 3.21 m3 ha-1 yr-1
Site Index 300 Index
Variable importance
LiD
AR
der
ived
SI
Av.
spr
ing
air t
emp.
Av.
spr
ing
sola
r rad
.E
VI
Band
5 M
E-2
5A
geN
IR re
fect
ance
Sim
ple
ratio
ND
VI
Ban
d 5
CO
R-1
5B
1 V
AR
-25
Band
4 M
E-2
5Bl
ue re
flect
ance
Red
refle
ctan
ceB
and
4 SD
-25
Gre
en re
flect
ance GR
Slo
pe
Mea
n Im
porta
nce
0
1
2
3
4
5
Variable included in the model
LiD
AR
der
ived
SI
Av.
spr
. air
tem
pBa
nd 5
ME
-25
Red
refle
ctan
ceN
IR re
fect
ance
Age
Blue
refle
ctan
ce EV
IBa
nd 5
EN
T-15
B
and
4 SD
-25
ND
VI
Sim
ple
ratio
Band
4 A
SM
-25
Band
5 C
OR
-15
Band
4 M
E-2
5A
v. s
pr. s
olar
rad
Ban
d 5
CO
N-1
5 A
v. s
olar
rad.
Mea
n F-
valu
e 0
300
600
900(a)
(b)
LiD
AR d
eriv
ed S
IAv
. spr
. air
tem
pAv
. sum
. VPD Age
Perc
. ret
. > m
ean
Can
opy
relie
f rat
ioN
DVI
NIR
refle
ctan
ce EVI
Brig
htne
ssB
and
5 M
E-2
5Le
ngth
and
slo
pe fa
ctor
Red
refle
ctan
ceG
reen
refle
ctan
ceIn
tLku
rtosi
sC
N ra
tioR
ed e
dge
ratio
Ban
d 5
EN
T-5
Mea
n Im
porta
nce
-1
0
1
2
3
4
Variable included in the model Li
DA
R d
eriv
ed S
IAv
. spr
. air
tem
pB
and
5 M
E-25
ND
VI
Can
opy
relie
f rat
ioE
VI
NIR
refe
ctan
ceAg
eP
erc.
ret.
> m
ean
Brig
htne
ssR
ed re
flect
ance
Band
4 E
NT-
25
Ban
d 2
VAR
- 25
Red
edg
e ra
tioG
reen
refle
ctan
ceA
v. s
olar
rad.
In
tLku
rtosi
sC
N ra
tio
Mea
n F-
valu
e
0
100
200
300
400(a)
(b)
Results – Top three variables
Metric Type VariablesSite Index P LiDAR derived Site Index; Av. spring air temp; NIR mean texture
NP LiDAR derived Site Index; Av. spring air temp; Av. spring solar rad.
300 Index P LiDAR derived Site Index; Av. spring air temp; NIR mean textureNP LiDAR derived Site Index; Av. spring air temp; Av. summer VPD
Research gains
Predicted Site Index (m)20 25 30 35 40
20
25
30
35
40
20
25
30
35
40
45
Predicted 300 Index (m3 ha-1 yr-1)10 15 20 25 30 35
10
15
20
25
30
35
Mea
sure
d Si
te In
dex
(m)
10
15
20
25
30
35
40
Mea
sure
d 30
0 In
dex
(m3 h
a-1 y
r-1)
Using Environmental Surfaces (previous method)
Using all variables, including LiDAR
Research gains
Predicted Site Index (m)20 25 30 35 40
20
25
30
35
40
20
25
30
35
40
45
Predicted 300 Index (m3 ha-1 yr-1)10 15 20 25 30 35
10
15
20
25
30
35
Mea
sure
d Si
te In
dex
(m)
10
15
20
25
30
35
40
Mea
sure
d 30
0 In
dex
(m3 h
a-1 y
r-1)
Using Environmental Surfaces (previous method)
Using all variables, including LiDAR
R2 = 0.57RMSE = 3.55 m3 ha-1 yr-1
R2 = 0.79RMSE = 2.45 m3 ha-1 yr-1
R2 = 0.76RMSE = 2.30 m
R2 = 0.91RMSE = 1.40 m
Summary
• Models of Site Index and 300 Index created
• Significant gains in precision possible using information from LiDAR and RapidEye
• For models without age, best data source were environmental surfaces + RapidEye
• For models with age, best data source was LiDAR
• RapidEye and environmental surfaces useful cost effective alternative to LiDAR
Acknowledgements
• GCFF programme jointly funded by MBIE and NZ Forest Growers Levy Trust
• Timberlands
• Professional contribution of all field teams involved
http://research.nzfoa.org.nz/www.scionresearch.com/gcff
Michael WattResearch Leader – Forest Operations and Monitoring
Date:24 March 2015