Combining SAR and optical time series for
monitoring tropical forest change (Group 4)
Johannes Reiche, Jan Verbesselt, Dirk Hoekman, Eliakim Hamunyela, Sytze de
Bruin, Arun Pratihast, Jan Pokorn, Christos Sotiropoulos, Martin Herold*
Wolf Forstreuter **
* Laboratory of Geo-Information Science and Remote Sensing, Wageningen University &
Research, The Netherlands
** SOPAC, Fiji
Content
1. Introduction
2. “Bayesian approach" to combine multi-sensor time series for
NRT deforestation detection
3. Work progress (Bolivia, Fiji, Ethiopia)
1
Introduction: NRT monitoring of tropical forest change
Near-real time (NRT) = capacity to detect changes in new satellite
images once they are available
Optical time series approaches: mainly for detecting historical
changes; limited in cloud covered regions
SAR time series approaches: few exist, but limited observations in
the past & commercial distribution
Multi-sensor SAR-optical: “pioneer” approaches exist, but limited
to combining Landsat & ALOS PALSAR (e.g. Lehmann et al., 2012,
Reiche et al. 2015a/b)
2
-> Need & potential for SAR-optical multi-sensor combination!
3
Introduction: Medium resolution optical and SAR sensors
Reiche et al., 2016 (Nature Climate Change, 6, 120-122): Combining satellite data for better tropical forest monitoring.
Introduction: Correlation and/or fusion of optical and SAR
time series not directly feasible!
90% missing data
Lan
dsa
t N
DV
I
PALS
AR
HV
HH
i. they are discrete
ii. their individual observations are non-equidistant in time
iii. their observation times are not identical
iv. their scales/unit are not directly compatible (e.g. NDVI vs. SAR backscatter)
4
Bayesian approach to combine multi-sensor
time series for NRT deforestation detection
Reiche et al., 2015, (Remote Sensing, 7, 4973-4996): A Bayesian Approach to Combine Landsat and ALOS PALSAR Time
Series for Near Real-Time Deforestation Detection.
5
Input: Multi-sensor time series observations in NRT environment
s1t-2
Past observations
s2t-1 s1t s1t+1 s2t+2 s2t+n
Current observation
Future observations
s2t+2
sensor 2 at time = t+2
(2nd future observation)
sensor 2 =
ALOS PALSAR HV
backscatter
Bayesian approach
s1t
sensor 1 at time = t
(current observation)
sensor 1 =
Landsat NDVI
6
Step 1: Deriving and combining TS of conditional non-forest (NF) probabilities
s1t-2
Past observations
s2t-1 s1t s1t+1 s2t+2 s2t+n
Current observation
Future observations
Sensor specific forest (F) and non-forest (NF) pdfs
sNFt-2 sNF
t-1 sNFt sNF
t+1 sNFt+2 sNF
t+i Conditional non-forest probabilities
Multi-sensor time series observations
Bayesian approach
7
Step 1: Deriving and combining TS of conditional non-forest (NF) probabilities
s1t-2
sNFt-2 NF
F
Sensor specific forest (F) and non-forest (NF) pdfs
Sensor specific forest (F) and non-forest (NF) pdfs for sensor 1 (Landsat NDVI)
1
( 1 | )( | 1 )
( 1 | ) ( 1 | )
tt s
t t
p s NFP NF s for t T
p s NF p s F
P(NF|s1) = conditional NF probability
p(s1|F) = conditional probability of s1 given the presence of F
p(s1|NF) = conditional probability of s1 given the presence of NF
1( | 1 ) :
( | ) :
t sNF
t
t sn
P NF s t T
s
P NF sn t T
sNF = combined time series of conditional NF probabilities
s1t-2 (NDVI ) = 0.7 = 0.1
= 0.18
= 0.64
Bayesian approach
8
Step 1: Deriving and combining TS of conditional non-forest (NF) probabilities
s1t-2
Past observations
s2t-1 s1t s1t+1 s2t+2 s2t+n
Current observation
Future observations
Sensor specific forest (F) and non-forest (NF) pdfs
sNFt-2 sNF
t-1 sNFt sNF
t+1 sNFt+2 sNF
t+i Conditional non-forest probabilities
Multi-sensor time series observations
Bayesian approach
9
s1t-2
Past observations
s2t-1 s1t s1t+1 s2t+2 s2t+n
Current observation
Future observations
sNFt-2 sNF
t-1 sNFt sNF
t+1 sNFt+2 sNF
t+i Conditional non-forest probabilities
Multi-sensor time series observations
Conditional probability of deforestation at t
( | )NFt t iP D s
Step 2: Iterative Bayesian updating of the probability of deforestation
Bayesian approach
10
s1t-2
Past observations
s2t-1 s1t
Current observation
sNFt-2 sNF
t-1 sNFt
Conditional probability of deforestation at t
( | )NFt t iP D s
• If conditional NF probability (sNF ) > 0.5 --> Flag potential deforestation --> Calculate conditional probability of deforestation, P(Dt|sNF
t+i)
Bayesian probability updating
1( | ) ( | )( | )
( )
NF NFNF t i t t i
t t i NFt i
P s D P D sP D s
P s
Step 2a: Flag potential changes and calculate probability of deforestation
Bayesian approach
13
s1t-2
Past observations
s2t-1 s1t
Current observation
sNFt-2 sNF
t-1 sNFt
Conditional probability of deforestation at t
( | )NFt t iP D s
• Future observations used as new evidence to update P(Dt|sNF
t+i) and to confirm or reject the deforestation event by exceeding a threshold
s1t+1
sNFt+1
Step 2b: Iterative Bayesian updating using upcoming observations
Bayesian approach
• If conditional NF probability (sNF ) > 0.5 --> Flag potential deforestation --> Calculate conditional probability of deforestation, P(Dt|sNF
t+i)
Bayesian probability updating
1( | ) ( | )( | )
( )
NF NFNF t i t t i
t t i NFt i
P s D P D sP D s
P s
1.0
0.8
0.6
0.4
-4
-5
-6
-7
ALO
S PA
LSA
R
H
VH
Hm
t [d
B]
Lan
dsa
t N
DV
I
2005.0 2006.0 2007.0 2008.0 2009.0 2010.0
remaining cloud
11
Original time series (top)
Bayesian approach: single-pixel example
1.0
0.8
0.6
0.4
1.0
0.8
0.6
0.4
0.2
0
-4
-5
-6
-7
ALO
S PA
LSA
R
H
VH
Hm
t [d
B]
Lan
dsa
t N
DV
I sN
F
2005.0 2006.0 2007.0 2008.0 2009.0 2010.0
old flagged change
TF (DOY: 2008.219)
T (DOY: 2008.266)
Reference: 2008.3 (DOY: 2008.182 – 2008.273)
remaining cloud
Original time series (top) Conditional NF probabilities (sNF) (bottom)
11
-4
-5
-6
-7
ALO
S PA
LSA
R
H
VH
Hm
t [d
B]
remaining cloud
old flagged change
TF (DOY: 2008.219)
T (DOY: 2008.266)
old flagged change
Change flagged (DOY: 2008.219)
Change confirmed (DOY: 2008.266)
Bayesian approach: single-pixel example
Work progress
1. Bolivia: NRT deforestation detection in dry forest
(Sentinel-1 + PALSAR-2 + Landsat)
2. Fiji (FIJ-1): Post-cyclone forest disturbance detection
(Sentinel-1 + PALSAR-2 + Landsat)
[MSC thesis of Jan Pokorn; started 09/16)
3. Ethiopia (ETH-1): Combining remote sensing and community-based
data streams
[MSC thesis Christos Sotiropoulos; started 10/16)
12
Size: 100 x 100 km
Tropical dry broadleaf
forests
Industrial logging
Frequent cloud cover
13
Combining Sentinel-1, PALSAR-2 & Landsat for NRT
deforestation detection in dry forest (Bolivia)
Sentinel-1 VV
Reiche et al., (in prep): Near real-time deforestation monitoring in dry tropical forest by combining Sentinel-1, PALSAR-2 and Landsat time series imagery.
Time series processing
Sentinel-1 VV & ALSO-2 PALSAR-2 HVHH-ratio: standard processing,
quality control, multi-temp filtering (Gamma software)
Landsat NDVI: standard processing, quality control (Ledaps, Fmask)
Co-registration of Sentinel-1, PALSAR-2 to Landsat (Gamma software)
Spatial normalisation to reduce seasonality in time series (Eliakim
Hamunyela et al., 2016, RSE)
15
Spatial normalisation
Apply forest mask
Image-wise spatial normalisation
Normalised pixel = pixel – 95th percentile of surrounding area
Landsat
Sentinel-1
16
Spatial normalisation
Landsat
16
PALSAR-2
Apply forest mask
Image-wise spatial normalisation
Normalised pixel = pixel – 95th percentile of surrounding area
Detected deforestation 10/2015 – 10/2016
17 0 10 km
Change flagged (DOY):2016.364 Change confirmed (DOY): 2016.5
2016/10
2015/10
Detecting post-cyclone forest disturbance (FIJ-1)
Cyclone Winston (20-02-2016)
● 40,000 homes were damaged or destroyed
● Fiji Pine lost 2 million dollars
Close cooperation with Fiji Forestry Department (FFD) & SOPAC
Reference data: VHR (Airbus), Ground data (FFD)
MSC thesis Jan Pokorn; started in 09/2016
18
Conclusion
Thanks GFOI, ESA & K&C!
Very, very exciting!
Combining SAR and optical time series makes a lot of sense
Work plan for next year:
Finish Bolivia work (validation & publishing)
Continue work in Fiji (FIJ-1) & Ethiopia (ETH-1)
Further develop “Bayesian approach” beyond “pioneer level” (e.g.
multi-modal and multi-spectral class descriptions)
R-package on “Bayesian approach” with reproducible examples
24
“More observations make everything better!”
Curtis Woodcock, ESA LPS 2016
Reiche et al., 2015, (Remote Sensing, 7, 4973-4996): A Bayesian Approach to Combine Landsat and ALOS PALSAR Time Series
for Near Real-Time Deforestation Detection.
Reiche et al., 2016 (Nature Climate Change, 6, 120-122): Combining satellite data for better tropical forest monitoring.
Reiche et al., (in prep): Near real-time deforestation monitoring in dry
tropical forest by combining Sentinel-1, PALSAR-2 and Landsat time series imagery.