Post on 30-Jul-2020
Anthony Vodacek
Digital Imaging and Remote Sensing
Laboratory (DIRS)
Center for Imaging Science
Rochester Institute of Technology
October 25, 2007
Assimilating remote sensing with dynamic
data-driven environmental models
Acknowledgments
Collaborators
Jan Mandel - UC-Denver - Data assimilation
Janice Coen - NCAR - Fire propagation modeling
Craig Douglas - UKY - DDDAS
Al Garrett - DOE/SRNL - Hydrodynamic modeling
Bob Kremens - RIT - Fire science and ground measurements
Students
Yan Li - Data assimilation and hydrodynamic modeling
Alvin Spivey - Hydrologic modeling/LCLU
Zhen Wang - Fire propagation modeling
Yushan Zhu - Aquatic optics
Staff
Scott Brown - DIRSIG Project Leader
Niek Sanders - DIRSIG developer
Jason Faulring - Airborne remote sensing
Funding
NSF - CNS-0324989 and CNS-0719626
NOAA - NA06OAR4600217
Dynamic Data-Driven Applications
Systems (DDDAS)
“DDDAS entails the ability to dynamically incorporate additional
data into an executing application, and in reverse, the ability of
an application to dynamically steer the measurement process”
www.dddas.org
Dynamic System
• System model
– Describes how the state of the system evolves over time
• Measurement model
– Describes how measurements are related to state variables - “synthetic data”
11 !! +=tttwAxx
state variable
(TSS concentration
or fire position)system model
(hydrodynamic model
or fire propagation model)
uncertainty in the system
tttvHx +=!
Measurement
(visible reflectance
or thermal radiance)
measurement model
(Hydrolight
or DIRSIG)
uncertainty in the measurement
MODEL
Synthetic
data or image
Real data
or images
output
data
assimilation
The synthetic data generation is an important aspect
and may or may not be complicated to produce
Ensemble
Kalman Filter
(EnKF)
update
Image Data: MODIS
Image Data: Landsat
Image Data: MISI
suspended sediment lots of algae
Lake Ontario
colored dissolved organic matter
very clear water
Water Quality Motivation:
Public Beach Closures
1. Swimming season 2003: June 21 ~ Sep. 1
2. Beach open 48 days (66%)
3. Closed 25 days (34%)
4. Genesee River
• Sediment dominated plume
• Wind and lake currents can push
polluted plume water onto the
beach
• Poor water clarity
• Prior day bacterial counts
• Local rainfall
• Excessive algae
2003 Ontario beach monitoring report. Monroe County Health Department, Bureau of Environmental Quality. September, 2003.
Water Quality Objective
1. Develop an integrated water quality modeling system to simulate thehydrodynamic and optical features in Lake Ontario and its coastalareas
2. Assimilate intermittent satellite or airborne data into a hydrodynamicmodel to obtain a better estimate of model state via an EnsembleKalman Filter
3. Evaluate the EnKF result at single point and spatially
4. Use the results to help forecast beach closures (funded to do this)
5. Use the knowledge of system uncertainty to prioritize or directmanagement efforts to the greatest effect and efficiency
Kalman Filter Notations
+post update
-prior to update
Covariance between
ensemble model states
Ptsystem uncertainty
EstimatedQmodel noise
Derived from MODISNtmeasurement noise
Remote-sensing reflectancemeasured observation
HydrolightHnon-linear measurement
model
ALGEAnon-linear system model
TSS concentrationXtstate variable
t!
+
!
!=
1ˆˆttxAx
11 !
+
!
!+= t
T
tt QAAPP
!
ˆ x t
+ = ˆ x t
" + Gt#
t"H( ˆ x
t
")[ ]
!
Pt
+= (I "G
tH( ˆ x
t
"))P
t
"
!
Gt
= Pt
"H
T( ˆ x
t
") H( ˆ x
t
")P
t
"H
T( ˆ x
t
") + N
t[ ]"1
Time Update (Prediction)
(1) Predict the state variable
(2) Predict the system uncertainty
Measurement Update (Correction)
(1) Compute the Kalman gain
(2) Update estimate with measurement
(3) Update the system uncertainty
Initial estimates for x and PWelch, G. and Bishop, G. (2006) An introduction to the Kalman Filter. Department of
Computer Science, University of North Carolina at Chapel Hill.
Kalman Filter Operation Cycle
Kalman Gain
• Kalman gain is an optimal weighting factor that minimizes
the a posteriori error covariance
• As the measurement noise N approaches zero,
• As the system uncertainty P approaches zero,
NHHP
HPG
T
T
+=
!
!
!
N"0limG =
P#H
T
HP#H
T=1
H
!
P"0limG = 0
large value
Weather variables
• Air/Dew point
temperature
• Wind speed/direction
• Cloud cover/height
• Precipitable water
• Pressure
BASINS
HSPF
3D TSS & CDOM
concentration
profiles
Modeled Rrs
EnKF
MODIS 250 m reflectance data
Updated TSS
concentration profiles
ALGE Hydrolight
• IOPs (a, b, bb)
• CHL concentration profiles
TSS & CDOM
concentration
and discharge
Modeling System Design Overview
Hydrodynamic model: ALGE
low
high
TSS
concentration(g/m3)
Synthetic image: Hydrolight
0
0.005
0.01
0.015
0.02
0.025
0.03
0.035
0.04
0.045
0.05
0 50 100 150 200 250 300
TSS Concentration (g/m3)
Ca
lcu
late
d R
rs (s
r-1)
0
0.05
0.1
0.15
0.2
0.25
0.3
300 400 500 600 700 800
Wavelength (nm)
Sp
ec
ific
ab
so
rpti
on
co
eff
icie
nt
(m2/m
g)
3D TSS concentration
profiles
Modeled Rrs
Hydrolight
• IOPs (a, b, bb)
• CHL and CDOM concentration
profiles0
0.02
0.04
0.06
0.08
375 475 575 675 775
Wavelength (nm)
Sp
ec
ific
ab
so
rpti
on
co
eff
icie
nt
(m2/m
g)
MODIS 250 m Surface Reflectance Data (645 nm)
(Aug. 18, 2003)
August 18, 2003 – beach closed
due to westward flowing plume
Observation Error Distribution for MODIS
• Error distribution is approximately Gaussian
0
2
4
6
8
10
12
-0
.00
16
-0
.00
14
-0
.00
12
-0
.00
10
-0
.00
08
-0
.00
06
-0
.00
04
-0
.00
02
0.0
00
0
0.0
00
2
0.0
00
4
0.0
00
6
0.0
00
8
0.0
01
0
0.0
01
2
0.0
01
4MODIS Reflectance Observation Error
Freq
uen
cy (
%)
Experimental Data
• TSS-dominated plume dissipation was modeled by ALGE
– Study area: Rochester Embayment
– Study period: July 26 ~ Aug. 18, 2003 (24 days)
– Nudged from whole lake simulation
– Bathymetry
Hydrodynamic simulation
grid spacing:
135 m x 135 m x 3 mGenesee River
depth m
0
2
4
6
8
10
7/25 7/29 8/2 8/6 8/10 8/14 8/18
Date
Win
d s
pe
ed
(m
/s)
0
50
100
150
200
250
300
350
7/25 7/29 8/2 8/6 8/10 8/14 8/18
Date
Dis
ch
arg
e (
m/s
3)
0
50
100
150
200
250
7/25 7/29 8/2 8/6 8/10 8/14 8/18
Date
Riv
er
TS
S c
on
ce
ntr
ati
on
(g
/m3)
-90
0
90
180
270
360
7/25 7/29 8/2 8/6 8/10 8/14 8/18
Date
Win
d d
irecti
on
(d
eg
ree)
Varying Forcing Variables
7/29
west ( ), 7.5mph8/13
southeast ( ), 3 mph
8/8
northwest ( ), 9 mph
7/30
southwest ( ), 4 mph
8/18
East ( ), 10 mph
8/16
Northwest ( ), 12 mph
8/15
Southwest ( ), 8 mph8/14
West ( ), 9 mph
False-color image (R: 858 nm G: 645 nm B: 645 nm)
Vegetation appears red, water appears dark, and plume appears blue
MODIS Reflectance Images showing river plume,
with beach closings and wind direction indicatedOpen
Closed
Closed
ClosedClosed
Closed
Open
Open
Single point EnKF estimate at river mouth
0
40
80
120
160
200
7/25 7/29 8/2 8/6 8/10 8/14 8/18
Date
TS
S c
on
cen
trati
on
(g
/m3)
truth
EnKF estimate
ensemble
model states
0
4
8
12
16
20
7/25 7/29 8/2 8/6 8/10 8/14 8/18
Date
Err
or
of
TS
S c
on
ce
ntr
ati
on
(g
/m3)
Single point EnKF estimate error
Red rectangles indicate a day
MODIS data is available for
updating
• Error of the EnKF estimates decreases with a measurement update
•The EnKF estimates deteriorate until new MODIS data become available
8/14 8/15 8/16 8/18
MODIS
Comparison of spatial estimate of plume
8/14 8/15 8/16 8/18
MODIS
Open-loop
Comparison of spatial estimate of plume
Comparison of spatial estimate of plume
8/14 8/15 8/16 8/18
EnKF estimate
Open-loop
MODIS
90
116
143
170
TSS
concentration (g/m3)
0.0035EnKF estimate
0.0083Ensemble mean
RMSE for all water pixels
(reflectance units)
Spatial EnKF error test for August 18, 2003
Maximum MODIS reflectance value = 0.045 reflectance units
Maximum EnKF reflectance value = 0.044 reflectance units
Future Water Quality Work
• Better weather data (lake breeze)
• Include the variability of model parameters and adjust the parameters based onthe assimilation results
– Particle density and diameter
– Some physical parameters
• Assimilate other satellite/airborne data and field observations into thehydrodynamic model as well
– MODIS Thermal (day/night)
– Landsat series
– WASP
• Increase the number of ensemble members to improve the estimation accuracy
• Batch filtering?
ITR/NGS: Collaborative Research: DDDAS: Data Dynamic Simulation
for Disaster Management
Movie by Janice Coen, NCAR
Fire images captured with the RIT WASP Lite sensor of a
prescribed burn in Kentucky in April 2007
panchromatic
thermal infrared
image segmentation with
Multivariate Gaussian Hidden
Markov Random Field
automated extraction
of fire line parameters
smoke
fire
burn
scar
AVIRIS
MODIS
airborne
simulator
Image processing methods for fire features
atmosphere-fire model
EnKF
ground sensors and
airborne imagerydata transmission
and visualization
synthetic senor and
image dataoutput
update
Fire Propagation Modeling System
Estimating the infrared scene radiance requires 3 phenomena to be modeled:
1. Radiation from the hot ground under the fire front and the cooling of the
ground after the fire front passes.
2. The direct radiation to the sensor from the 3D flame.
3. The radiation from the 3D flame that is reflected from the nearby ground.
Generating the synthetic image
16 minutes after ignition 74 minutes after ignition
2D temperature maps with cooling
Flame tilt
Side projections of 3D flame structure
estimated from model output heat flux
t=0s t=12s t=24s t=36s
DIRSIG: The Digital Imaging and Remote Sensing Image Generation ModelA physics-based ray tracing model for generating synthetic remote sensing scenes
near infrared simulation of a section of the RIT campus
0 0.108 W cm-2 sr-1 µm-1
Final rendering: synthetic shortwave infrared scene of a grassfire at night
0 0.459 W cm-2 sr-1 µm-1
Final rendering: synthetic midwave infrared scene of a grassfire at night
0 0.117 W cm-2 sr-1 µm-1
Final rendering: synthetic longwave infrared scene of a grassfire at night
Short wave infrared
Mid wave infrared
Wildfire Airborne Sensor Project images
of a wildfire in Montana