1) The EO-1 Hyperion instrument has collected over 65,000 scenes over its 12-year mission to study land and coastal ecosystems using imaging spectroscopy.
2) Studies using Hyperion data have identified spectral indices related to chlorophyll that correlate with carbon flux measurements at different sites, including a Zambian woodland and North Carolina forest sites.
3) Time series of Hyperion data at flux tower sites show seasonal changes in these spectral indices that match patterns in ecosystem carbon uptake and release.
EO-1/HYPERION: NEARING TWELVE YEARS OF SUCCESSFUL MISSION SCIENCE OPERATION AND FUTURE PLANS
1. EO-1/HYPERION:
NEARING TWELVE YEARS OF
SUCCESSFUL MISSION SCIENCE
OPERATION AND FUTURE PLANS
Elizabeth M. Middleton
NASA/Goddard Space Flight Center, USA
Petya K. E. Campbell1, K. Fred Huemmrich1, Qingyuan Zhang2, Yen-Ben Cheng3, David
Landis4, Stephen Ungar2, Lawrence Ong5, and Nathan Pollack5
1 University of Maryland Baltimore County
2 Universities Space Research Association (USRA)
3 Earth Resources Technology, Inc.
4 Sigma Space Corp.
5 Science Systems and Applications, Inc.
IGARSS’12
MO3.3 Spaceborne Imaging Spectroscopy Missions:
Updates, and Global Datasets and Products [#4254]
Munich, Germany , July 23, 2012
2. Overview of the EO-1 Mission
Science Office Activities
Hyperion
• Acquisitions and Data Quality Checks
• Support New Algorithms (fAPARchl, PRI)
• Conduct Field Tests
• Comparisons with MODIS results
• Conduct Comparisons with Flux Towers
4. N A. B.
Time Series for CEOS Cal/Val Sites
Temporal variation in spectral characteristics, Railroad
Valley, NV Similar datasets are being assembled at other CEOS
Cal/Val and LPV sites + +
N
4 km
-1 -0.5 0 0.5 +1
Railroad Valley Playa site (cross):
A. Natural color composite (RGB:
651,549,447), B. Getis Gi
statistics, displaying the
homogeneous regions
Mean reflectance spectra (solid line)
Campbell et al. 2012 Standard deviations (dashed blue line)
5. EO-1 Hyperion Image Processing
Level 1R Hyperion data were atmospherically corrected using the
Atmosphere CORrection Now (ACORN) model.
Reflectance spectra were extracted in the vicinity of the existing flux
towers, from 30-50 pixels depending on the site size.
700 ice AC ice AT 600
bright target AC bright target AT corn (r = 0.95)
550
600 corn AC corn AT forest (r = 0.98)
lichens AC lichens AT 500 water (r = 0.92)
forest AC fores AT 450
500 bright target (r = 0.97)
water AC water AT
400 lichens (r = 0.98)
350 ice (r = 0.99)
400
300
300 250
O
N
A
R
C 200
%
n
aR
e
lc
t
)
(f
200 150
100
100 50
0
0
-50
450 700 950 1200 1450 1700 1950 2200 2450 -50 0 50 100 150 200 250 300 350 400 450 500 550 600
Wavelength (nm) ATREM
6. Scaling Fluxes to Aircraft
Imagery of cornfield from Airborne Imaging Spectrometer for Applications
(AISA) data collected on September 14, 2009. Left panel shows fAPAR from
NDVI; middle panel is PRI; and right panel is modeled GEP in mg CO2 m-2
s-1 using the model derived from ground reflectance data.
7. USDA Cornfield site
2008
EO-1 Hyperion
True color
fAPARcanopy
DOY 108 172 190 195 231 277 2008
Spring Summer Fall
8. fAPARcanopy
fAPARleaf
fAPARchl
fAPARNPV
DOY 108 172 190 195 231 277 2008
Spring Summer Fall
9. 2.5 LUEchl and PRI: in situ ASD canopy measurements
0.03
0.02
2 PRI= (ρ 531-ρ 570)/(ρ 531+ρ 570) 0.01
0
1.5 -0.01
LUEchl
-0.02
pri
1 -0.03
-0.04
0.5 -0.05
In situ canopy and leaf measurement dates
-0.06
0 -0.07
07/13/08 07/23/08 08/02/08 08/12/08 08/22/08 09/01/08 09/11/08 09/21/08 10/01/08 10/11/08
2.5
2
y = 23.969x + 1.8647
LUEchl(g mol-1)
2
LUEchl vs. PRI R = 0.8306
1.5
y = 23.97x + 1.86
1
r2 = 0.83
0.5
0
-0.07 -0.06 -0.05 -0.04 -0.03 -0.02 -0.01 0 0.01 0.02
PRI
10. USDA/ OPE3 Corn Field
Compare LUEchl vs. PRI: Hyperion [▲] and in situ ASD measurements [ ]
PRI= (ρ 531-ρ 570)/(ρ 531+ρ 570)
30 m, 10 nm bands Hyperion = ▲
Triangles over Circles are for the 5 days having both ASD and Hyperion images (2008
DOY 172, 190, 195, 231, 277). Hyperion data: 30 m, 10 nm bands.
11. Product Prototyping for HyspIRI
Comparisons of GEP from various algorithms
0 14.34 0 6.74 0 4.85
gCm-2d-1 gCm-2d-1 gCm-2d-1
60m Hyperion 60m Hyperion 60m simulated MOD17 1km GPP
RGB PRI & fAPARchl MOD17 Cheng et al. 2011. HyspIRI Symposium
12. Product Prototyping for HyspIRI
Comparisons of GEP from various algorithms
12
10
-2 -1
8
d )
6
4
m
G
P
C
E
g
(
2
0
OPE3 flux PRI MOD17 MOD17
tower fAPARchl mockup GPP
0 14.34 0 6.74 0 4.85
gCm-2d-1 gCm-2d-1 gCm-2d-1
60m Hyperion 60m Hyperion 60m simulated MOD17 1km GPP
RGB PRI & fAPARchl MOD17 Cheng et al. 2011. HyspIRI Symposium
13. USDA/ Beltsville Field
MAIAC-MODIS fAPARchl and PRI (488)
4.5
LUEchl (g mol-1)
4 y = -29.291x + 7.1335
3.5
3 R2 = 0.7647
2.5
2
1.5
1
0.5
0
-0.5
0.1 0.12 0.14 0.16 0.18 0.2 0.22 0.24 0.26
PRI (488)
14. MODIS based fAPARchl and PRI (488)
@ Great White Mountain flux tower site, China
8
y = -19.411x + 7.5556
7
R2 = 0.7841
6
LUEchl (g MJ-1)
5
4
3
2
1
0
0 0.1 0.2 0.3 0.4
PRI(488)
15. Scaling Light Use Efficiency in Arctic Tundra
From plot to region
- Plot level LUE
Chamber measurements of
photosynthesis of pure patches of
vascular plants, mosses, and
lichens
Spectral reflectance collected and
convolved to Hyperion bands
All observations from late July and
early August near Barrow, AK
- near peak of growing season
Data salvaged from old field work
16. Hyperion - Reflectance, Functional Type Cover, and LUE
Day 201, 2009, Image subset around Barrow, AK
Field measurements scaled to region find a 5-fold variation in LUE
R = Reflectance at 834 nm R = Vascular Plant Cover Light Use Efficiency
G = Reflectance at 671 nm G = Moss Cover (x10,000)
B = Reflectance at 549 nm B = Lichen Cover Based on coverage
Scale from 0 – 100%
17. Estimating Fluxes from MODIS Ocean Bands
in Canadian Forests
Examine Relationship between GEP and PRI*APAR from MODIS
- Mid-growing season data for 6 different forest types
- Fluxes from flux tower for time of overpass
- Distinct differences in responses among sites
18. Remote Sensing of Fluxes: Hyperion and Fluxnet
Can a single algorithm driven by hyperspectral satellite data
provide an estimate of carbon flux variables over a wide
range of sites?
Method: Matched flux data from LaThuile Fluxnet Synthesis
with Hyperion imagery
Standardized flux calculation for all sites
80 observations of 33 different
flux tower sites
Data from 2001 to 2007
Observed during mid-growing
season
Multiple vegetation types
Time Series at Flux Sites
La Thuile Flux Sites
CEOS Calibration Sites 18
19. CO2 Flux Data Processing
• Net Ecosystem Production (NEP, µmol m-2 s-1) is
the CO2 absorbed by the vegetation, measured
by the flux tower.
• Ecosystem Respiration (Reco) was calculated
from relationships developed between
nighttime Net Ecosystem Exchange (NEE) and air
temperature (sometimes, also soil moisture).
• Gross Ecosystem Production (GEP) is calculated
from the observed NEE and Reco.
21. Multi-Site Vegetation Index and LUE
• Best index (out of 107 tried) for overpass LUE was the first derivative at
732 nm divided by the derivative at 702 nm
79 Points 21
22. Multi-Site Vegetation Index and LUE
• Best index (out of 107 tried) for both overpass and daily LUE was the first
derivative at 732 nm divided by the derivative at 702 nm: D732/D702
At overpass time With daily fluxes
N =79
23. Stepwise Regression Test
• A wide range of bands can be used to produce good results (r > 0.82)
• Different input datasets chose different band sets for Daily LUE
- 38 different bands chosen in 11 runs (10 subsets and all points together)
- 9 runs chose band 732nm, 8 runs chose band 783nm
67 Points
24. Stepwise Linear Regression - LUE
• Circled points are outliers. R and RMSE calculated with outliers removed
79 Points
Used Bands: R569, R732, R742, R2093, R2133, R2153, R2375 Used Bands: R518, R539, R549, R732, R783, R915, R1023
25. Multi-Site Vegetation Index and Reco
• Best index (out of 107 tried) for Reco at overpass time was the
Normalized Difference Water Index (NDWI), using reflectances at 876
and 1245 nm. Reco = Ecosystem Respiration.
80 Points
26. Partial Least Squares –LUE at Overpass
• An example of an approach that utilizes all of the spectral information
red - PLS Weighting Factors
black - sample reflectance spectra
79 Points 26
27. Partial Least Squares – Reco at Overpass
red - PLS coefficients
black - sample reflectance spectra
79 Points
28. Results-Conclusions
• A common (global) spectral approach appears feasible. To derive it
we need:
– the capability of collecting hyperspectral observations of
globally-distributed sites representing a variety of vegetation
types
– the ability to make repeated measurements of each site
– Hyperion on EO-1 can provide data for these studies
• The strongest relationships use continuous spectra, narrow
wavelength bands, and/or derivative parameters
• Multiple algorithms and/or band combinations are effective
29. EO-1 Hyperion: Three Ecosystem Studies
Time Series
FLUX Site Locatio Climate Vegetation
Name
1. Mongu n
Zambia Temperate/ warm Kalahari/
summer Miombo
Woodland
2. Duke North Temperate/ no Hardwood
Carolina dry season/ hot forest/ Loblolly
USA summer pine
3. Konza Prairie Kansas Cold/ no dry Grassland
USA season/ hot
summer
Mongu
3.
2.
1.
MSO Sites
31. Bio-indicator Bands (nm) R2 [NEP (GEP)]
G32 R750, 700, 450 0.83 (0.81) NL
Dmax D max (650…750 nm) 0.77 (0.87) NL
Dmax / D704 D(690-730) 0.79 (0.80) NL
mND705 R750, 704, 450 0.75 (0.79) NL
RE1 Av. R 675…705 0.71 (0.56) NL
EVI R (NIR, Red, Blue) 0.73 (0.88) L
NDVI Av. R760-900, R620-690 0.52 (0.60) NL
G32, Associated with
Chlorophyll
(Gitelson et al. 2003)
32. Hyperion Spectral Indices and GEP at Mongu
B. Wet season (DOY 22) A. Dry season
(DOY 214)
DOY
The spectral bio-indicator associated with chlorophyll content (G32, green line) best
captured the CO2 dynamics related to vegetation phenology.
33. Mongu: Seasonal change in G32 & NEP
A. Dry season (DOY 214) G32 Estimated NEP (μmol m-2 s-1)
0 8 0 12
B. Wet season (DOY 22)
34. Duke, NC Loblolly Pine
DOY
Pine site
4000 DOY
3500
Mixed 6
Hardwood site Hardwoods
34
162
3000
180
2500 203
290
2000 300
1500
1000
500
0
450 700 950 1200 1450 1700 1950 2200 2450
35. Duke Forest : PRI4 & NEP
A. Winter (DOY 34) PRI4 NEP (μmol m-2 s-1)
LP
HW
B. Summer (DOY 203) -3 0 28
3
LP
HW
36. Bio- indicators of Photosynthetic Function
Loblolly Pine (LP)
Index Bands (nm) R2 [NEP (GEP) LUE]
PRI1 531, 570 0.84 (0.73) L
PRI4 531, 670 0.75 (0.63) 0.73 L
DPI D 680, 710, 690 0.91 (0.44) NL
NDWI 870, 1240 0.76 (0.60) L
NDVI NIR, Red 0.19 (0.48) L
Hardwoods (HW)
Index Bands (nm) R2 [NEP (GEP) LUE]
PRI4 531, 670 0.84 (0.48) NL
Dmax D max (650…750 nm) 0.83 (0.40) NL
NDII 820, 1650 0.79 (0.34) L
EVI NIR, Red, Blue 0.84 (0.41) L
NDVI NIR, Red 0.63 (0.19) L
38. Normalized Difference Water Index
Konza (K), Mongu (M), Duke (D)
0.10
0.05
0.00
-0.05
Mongu
-0.10 Duke
W
D
N
I
-0.15 Konza
-0.20 y = -0.0002x2 + 0.0119x - 0.1395
R² = 0.74
-0.25
-5 0 5 10 15 20 25 30 35 40
NEP
39. All Towers: Midday GEP vs. APAR
75
Mongu Av. LUE = 0.011 mol/mol
65
Duke Av. LUE = 0.017 mol/mol
)
-1
s
55 Konza Av. LUE = 0.043 mol/mol
-2
45
y = 0.0166x - 0.2254
R² = 0.76
35
m
µ
o
(
l
y = 0.0428x - 11.256
25 R² = 0.92
15
M
G
d
P
a
E
y
y = 0.0106x + 1.728
i
5 R² = 0.85
-5
0 500 1000 1500 2000
Midday APAR (µmol m-2 s-1)
44. EO-1 Hyperion Spectral Bio-Indicator of GEP/NEP
Best Correlation to CO2 Uptake for Multiple Flux Sites
†
NEP – net ecosystem production, GEP – gross ecosystem production
‡
L – linear, NL – non-linear
Campbell et al. 2012
2012 William Nordberg Award
44
45. Findings
• In 3 vastly different ecosystems, continuous reflectance
data and a variety of spectral parameters, were
correlated well to CO2 flux parameters (e.g. NEP, GEE,
etc.). Imaging spectrometry provides spatial distribution
maps of CO2 fluxes absorbed by the vegetation.
• The bio-indicators with strongest relationships were
calculated using continuous spectra, using numerous
wavelengths associated with chlorophyll content and/or
derivative parameters.
• Common (global) spectral approach to trace vegetation
function and estimate it’s CO2 sequestration ability is
feasible. It requires:
– a diverse spectral coverage, representative of the major ecosystem types,
– spectral time series, to cover the dynamics within a cover type.
46. Remote Sensing of Fluxes
Hyperion and Flux Towers
• Hyperion on EO-1 provides us with two important capabilities:
– the capability of collecting hyperspectral observations of
globally-distributed sites, and
– the ability to make repeated measurements of a site
• Provides a dataset for testing and developing algorithms for global
data products
• The strongest relationships with carbon uptake parameters used
continuous spectra, numerous wavelengths associated with
chlorophyll content, and/or derivative parameters.
• A common (global) spectral approach appears feasible. To derive it
will require:
– Diverse coverage, representing major ecosystem types, and
– time series, to cover the dynamics within a cover type.
47. Recommendations
These studies utilize data from the existing flux
tower network
For many HyspIRI products we will need more
studies applying algorithms for a number of
different landcover types
- Use ground, aircraft, and satellite spectral reflectance
data
- Need to develop protocols for ground measurements
of potential HyspIRI products
- Need to establish network of sites measuring these
products
- These sites can grow into a HyspIRI cal/val network
48. EO-1 Future Plan
• Present Matsu Compute Cloud functionality
– Hyperion and ALI Level 1R processing
– Hyperion and ALI Level 1 G processing
– Web Coverage Processing Service (WCPS) – web service to rapidly create and
execute new algorithms for ALI and Hyperion data and includes:
• Atmospheric Correction
• ALI Pan Sharpening
• Flood water classifier for ALI
– Namibia Flood Dashboard (mashup of ground and multiple satellite data and data
products for floods)
• Augment the Matsu Cloud
- Automated co-registration of Hyperion (depending on funding availability)
-Tile cutouts for Hyperion
• Lunar Calibration Schemes
• Intelligent Payload Module
- High speed onboard processing for low latency products (target HyspIRI)
- Hyperion L0, L2 to emulate future HyspIRI data
- WCPS - upload algorithms in realtime to customize processing of EO-1 like data
- Core Flight Executive (cFE)
- CASPER – onboard planner used on EO-1 is part of testbed
49. Future Directions
• Expand the tests over additional ecosystem types (rain forest, temperate
and sub-arctic vegetation types);
• Test additional spectral approaches (e. g. feature depth analysis)
• Special Issue of IEEE JSTARS on EO-1 (Guest Ed., E.M. Middleton),
early 2014.
?
?
?
?
?
Editor's Notes
Getis Gi statistics, calculated using ENVI for a moving 3x3 windows (9 pixels, ~ 100 m), to permit visualization of the homogeneous regions. A cluster of pixels with high digital counts is indicated by largely positive Gi* values, while a cluster of pixels with low digital counts is indicated by largely negative Gi* values.
Cornfield: 0.61 0.29 0.30 0.34 0.82 0.74 Surrounding forest: 0.50 0.86 0.79 0.78 0.79 0.85 The cornfield site: we have 6 Hyperion images in 2008 from Spring, to Summer, to Fall. The first row are true color images. The second row are the fAPARcanopy images based on an empirical relationship between fAPARcanopy and NDVI, which is widely used. The cornfield is surrounded by forests. Corn was planted before day of year 195. The corn fAPARcanopy was saturated during its peak of growing season. For the surrounding forest, its fAPARcanopy saturated from late spring to summer.
Cornfield: 0.51 0.13 0.16 0.23 0.77 0.69 0.40 0.10 0.12 0.18 0.71 0.46 0.11 0.03 0.04 0.04 0.05 0.23 Surrounding forest: 0.22 0.82 0.80 0.80 0.80 0.71 0.15 0.62 0.51 0.50 0.48 0.50 0.06 0.20 0.29 0.30 0.32 0.21 Four rows: fAPARcanopy, fAPARleaf, fAPARchl and fAPARNPV. You can see how different these 3 fAPARs are from fAPARcanopy. fAPARleaf also has some saturation issue. The cornfield and the surrounding forest have distinct fAPARchl and fAPARNPV seasonality. Before corn was planted, both fAPARchl and fAPARNPV were low. Corn was the greenest on day 231 and fAPARNPV was low. During the senescence period, fAPARchl decreased and fAPARNPV increased. The forest was the greenest on day 172. after that, fAPARchl gradually decreased and fAPARNPV increased. fAPARNPV on days 190 -231 kept same.
For the MODIS observations over the forest site, here is the comparison between Light Use efficiency at chlorphyll level and the narrow band MODIS PRI. The R2 is 0.78, very nice correlation.
This raises issues that we will explain using the leaf/canopy models
R dynamics at Mongu are a result of biophysical stress caused by the seasonal rainfall, which is reinforced by the CO 2 flux.