Rangeland Carbon Fluxes in the Northern Great Plains
Wylie, B.K., T.G. Gilmanov, A.B. Frank, J.A. Morgan, M.R. Haferkamp, T.P. Meyers, E.A. Fosnight, L. Zhang
US Geological SurveyNational Center for Earth Resources Observation Science (EROS)
• USGS: Earth Surface Dynamics, Land Remote Sensing, and Geographic Analysis and Monitoring
• National Science Foundation/EPSCoR Grant No. 0091948 through the South Dakota Center for Biocomplexity Studies and by the State of South Dakota
• USAID GL-CRSP project Grant No. PCE-G-00-98-00036-00 to Univ. of California, Davis• USDA Agricultural Research Service Agriflux network• US Agency for International Development Grant No. PCE-G-98-00036-00, Global Bureau,
Office of Agriculture and Food Security
U.S. Department of the InteriorU.S. Geological Survey
Objectives
Exhaustively model and map Carbon fluxes in rangelands to quantify the fluxes and Investigate temporal and patterns to understand climate and environmental influences.
If NDVI < 0.40C fluxes = f(X1, X2, X4)
If NDVI >= 0.40 C fluxes = f(X1, X4, X6)
Localized Detailed Measurements
Spatial Variables Through Time
C Flux Mapping Model Tree
C Flux Maps
METHODS: Develop Empirical Model Tree Algorithms to Predict Flux Tower Carbon Observations From Spatial Data
Northern Great Plains
Lethbridge (LE), Canada: Larry Flanagan
Ft. Peck (FP), MT; Tilden Meyers
Mandan (MA), ND; Al Frank
Cheyenne (CN), WY: Jack Morgan
Miles City (MC), MT: Marshall Haferkamp
Cottonwood, SD
Brookings, SD
Flux Tower Data is used to Develop Models for the Mapping of 10-day Carbon Dynamics on Rangelands
New Towers
METHODS: Tower C Fluxes Associated with Basic Ecosystem Functions of Gross Primary Production (Pg) and Total Ecosystem
Respiration (Re) are Derived From Light Response Curves on 30 min. Datasets.
Tagir Gilmanov, SDSU, coordinator of the FLUXNET WORLDGRASSFLUX working group
Derives estimates of daytime respiration
independentindependent of nighttime flux data
(others often use relationships derived from nighttime fluxes and
apply them to daytime observations)
PROVIDES:1) Tower Pg, Re, and NEE2) Facilitates flux tower data gap filling3) Quantifies ecophysiological parameters
(quantum yield (α), maximum gross photosynthesis (β), & daytime respiration (γ))
P(Q,Ts ; α,β,γ 0,θ,κ )=
12θ
αQ + β − (αQ + β )2 − 4αβθQ⎛ ⎝
⎞ ⎠ − γ0 eκTs ,
NEE
-2
-1
0
1
2
3
4
5
6-Dec
25-Jan
15-Mar
4-May
23-Jun
12-Aug
1-Oct
20-Nov
9-Jan
g C
m-2
day
-1
NEEPgRe
Ecosystem Process Carbon Partitioning: Added Value
Bkwd_finalGPP_RE_NEEmodel4c.xls
Sou
rce
S
ink
Flux Tower Pg Group: Significant Early Rain,
High Pg0
1
2
3
4
5
6
7
8
1-Jan 20-Feb 10-Apr 30-May 19-Jul 7-Sep 27-Oct 16-Dec
Pg
(g C
m-2
/day
-1)
0
20
40
60
80
100
120
140
10-d
ay p
reci
pita
tion
(mm
)
CH 1998 PgMA 1999 PgCH 1998 pptMA 1999 ppt
Flux Tower Pg Group: Moderate Rain,
Moderate Pg
0
1
2
3
4
5
6
7
8
1-Jan 20-Feb 10-Apr 30-May 19-Jul 7-Sep 27-Oct 16-Dec
Day of Year
Pg (g
C m
-2/d
ay-1
)
0
20
40
60
80
100
120
140
10-d
ay p
reci
pita
tion
(mm
)
LE 2000 PgLE 2001 PgMC 2000 PgMC 2000 pptLE 2000 pptLE 2001 ppt
Flux Tower Pg Group: Low Rain, Low Pg
0
1
2
3
4
5
6
7
8
1-Jan 20-Feb 10-Apr 30-May 19-Jul 7-Sep 27-Oct 16-Dec
Day of Year
Pg
(g C
m-2
/day
-1)
0
20
40
60
80
100
120
140
10-d
ay p
reci
pita
tion
(mm
) FP 2000 PgMA 2000 PgMA 2001 PgMC 2001 PgFP 2000 pptMA 2000 pptMA 2001 pptMC 2001 ppt
Model Development Data Set Consists of Flux Tower Pg, Re, and NEE Combined With Spatial Variables from Remote Sensing and GIS
(10-day Time Step)
Variable name DefinitionSeasonally Dynamic
Solar Radiation function of DOY and latitudend Smoothed NDVIppt precipitationtemp temperaturepar PAR
Seasonally Dynamic but dependent on timesinceSost Days since Start of Season (NDVI metric)DOY day of year = photoperiodismsolstice days prior or post summer solstice
Annually Dynamicsosn NDVI at Start of Seasonsost Start of Season DOYtin Time Integrated NDVIyear Categorical variable for testing year effects
Site Dynamicsitecat Categorical variable for testing site effectsc4pct % C4 from STATSGOco3surf Surface calcium carbonates from STATSGOpctclaysurf % clay in surface layer from STATSGOpctgrass % grass from STATSGOomernick Categorical variable for testing ecoregion effectsElevationAspect
Tower versus Spatial Data- Takes into account the accuracy of spatial data sets when developing model.- Process-based models typically rely heavily on precipitation - Precipitation can have high spatial and temporal variability that is difficult to map accurately.
Precipitation
y = 1.0331x + 2.3167R2 = 0.4247
0
10
20
30
40
50
60
70
80
90
0 10 20 30 40 50 60 70 80 90
Mapped Precipitation (mm)
Tow
er P
reci
pita
tion
(mm
) .
Temperature
y = 1.0171x + 0.2487R2 = 0.9852
-10
0
10
20
30
-10 0 10 20 30
Mapped Temperature (oC)
Tow
er T
empe
ratu
re (
o C)
Distinct Model Trees
• Models that predict:– Net Ecosystem Exchange (NEE)– Gross Primary Production (Pg)– Respiration (Re)
Empirical Piecewise Regression Model
• Cubist– Parsimonious and transparent selection of variables– Stratification of data into homogeneous information
spaces– Multiple regression defined for each information
space
NEE Model example (g CO2 m-2 day-1)
Rule 1: [98 cases, mean -1.1538022, range -6.22692 to 5.404624, est err 1.2296274]
if nd <= 140 and sinceSost > 85then
NEE = 6.25352 + 0.138 nd - 0.091 ppt - 0.19 sosn - 0.009 par- 0.01 temp - 0.0015 sinceSost
Rule 2: [21 cases, mean 0.7320258, range -3.83196 to 4.013414, est err 1.8628249]
if nd > 140 and par <= 100then
NEE = 1.82742 + 0.046 par - 0.04 temp + 0.033 nd - 0.03 sosn- 0.008 ppt - 0.0014 sinceSost
Rule 3: [60 cases, mean 0.9998901, range -4.301141 to 7.026846, est err 1.7766731]
if nd <= 140 and sinceSost <= 85then
NEE = 3.42815 + 0.051 par - 0.046 temp - 0.054 tin + 0.043 nd- 0.04 sosn - 0.012 ppt - 0.0019 sinceSost
Flux Tower Training Data Set and Respective NEE Rules
-2
-1
0
1
2
3
4
6-Dec25-Jan15-Mar4-May23-Jun12-Aug1-Oct20-Nov9-Jan28-Feb
Date
Tow
er N
EE
(g C
m-2
day
-1)
rule 1, Rel. Err. 0.11rule 2, Rel. Err. 0.24 rule 3 Rel. Err. 0.16rule 4 Rel. Err. 0.25rule 5 Rel. Err. 0.22April - Oct.
Rules or respective “information spaces”
Theme Count Pct. Weight Pct.---------------------------------------------
par 429 34.3 1.712 13.2nd 390 31.2 2.981 22.9
sinceSost 379 30.3 0.732 5.6temp 40 3.2 3.360 25.8ppt 11 0.9 1.636 12.6tin 0 0.0 1.385 10.7sosn 0 0.0 1.193 9.2
Variables used to predict NEE
Theme Count Pct. Weight Pct.---------------------------------------------
nd 683 37.8 7.031 33.5par 508 28.1 5.322 25.3doy 177 9.8 2.231 10.6temp 136 7.5 2.222 10.6
sinceSost 132 7.3 1.437 6.8c4pct 111 6.1 0.448 2.1sost 60 3.3 0.827 3.9tin 0 0.0 1.482 7.1
Variables used to predict Pg
Variables used to predict Re
Theme Count Pct. Weight Pct.---------------------------------------------
temp 2848 40.4 3.540 4.0 nd 2244 31.8 22.192 25.2ppt 991 14.1 8.395 9.5par 343 4.9 10.789 12.3
sinceSost 284 4.0 12.412 14.1 doy 255 3.6 10.735 12.2
pctgrass 81 1.1 14.612 16.6tin 0 0.0 5.326 6.1
Model Accuracy Assessment
1/n | observed − predicted |∑MAD =
Range = 95th percentile - 5th percentile
Relative Error = MAD / Range
NGP winter NEE
-1.8
-1.6
-1.4
-1.2
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
-20 -15 -10 -5 0 5 10
Temperature (C)
NEE
(g C
m-2
day
-1)
Winter fluxes difficult to estimate….
30 min. data sets indicate snow depth and wind speed important
Winter fluxes (Nov. - Mar.) are need toconvert model tree maps (Apr. - Oct.) to Annual Fluxes
-0.8
-0.7
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0Mandan 01-02 Mandan 96-00 Miles City Lacombe Average
Winter C flux Grassland Studies
Ave
rage
NE
E (g
C m
-2 d
ay-1
)
Mandan 01-02, Gilmanov 2002; Mandan 96-00, Frank et al. 2002, Miles City, Gilmanov 2002; Lacombe, Baron et al. 2004
Cross validation
Five-fold NEE (g C m-2 day-1) cross-validation (random)
• Methodology to quantify robustness of small training data sets• Random cross-validation produces realistic error estimates• Cross-validation by sites evaluates influence of individual sites• Cross-validation by year evaluates influence of individual years
Average MAD = 0.48
Relative Error
0.6000.380.480.430.49Mean Absolute Difference
4444434343Hold out data sample size
Fold 5Fold 4Fold 3Fold 2Fold 1Random sample as test
Relative Error
0.6000.380.480.430.49Mean Absolute Difference
4444434343Hold out data sample size
Fold 5Fold 4Fold 3Fold 2Fold 1Random sample as test
Results
0.150.160.290.17Relative Error (test)0.480.490.910.55Mean Absolute Difference (test)
074972521Sample size (test)0.110.120.140.100.12Relative Error (train)0.350.390.440.330.37Mean Absolute Difference (train)
217143120192196Sample size (train)
None2001200019991998Year withheld as test
0.150.160.290.17Relative Error (test)0.480.490.910.55Mean Absolute Difference (test)
074972521Sample size (test)0.110.120.140.100.12Relative Error (train)0.350.390.440.330.37Mean Absolute Difference (train)
217143120192196Sample size (train)
None2001200019991998Year withheld as test
Cross-validation NEE (g C m-2 day-1) by site
Cross-validation NEE (g C m-2 day-1) by years0.180.170.100.210.14Relative Error0.590.550.310.660.46Mean Absolute Difference
04421217358Sample size (test)0.110.120.120.110.090.14Relative Error (train)
0.350.380.370.360.290.45Mean Absolute Difference (train)
217173196196144159Sample size (train)NoneMCCNFPMALESite withheld as test
0.180.170.100.210.14Relative Error0.590.550.310.660.46Mean Absolute Difference
04421217358Sample size (test)0.110.120.120.110.090.14Relative Error (train)
0.350.380.370.360.290.45Mean Absolute Difference (train)
217173196196144159Sample size (train)NoneMCCNFPMALESite withheld as test
Pg: Jack_site_summary3.xls
Robustness of Pg model tree: Cross Validation using Sites
• Cheyenne and Mandan are influential sites (high Pg in CH 1998 and MA 1999)(MA is only high precipitation, eastern site)
•Relative Errors on the other sites was from 7 to 12 %
Site Withheld CN FP LE MA MC NoneMAD (g C m-2 day-1) 0.94 0.32 0.46 1.65 0.53 0.32Relative Error 0.21 0.07 0.10 0.37 0.12 0.07N 21 21 58 73 43 217
NGP Modeled April - Oct. average Re (g C m-2 day-1) versus Observed Re
y = 0.8382x + 0.3206R2 = 0.639
1
1.5
2
2.5
1 1.5 2 2.5
Modeled Re
Tow
er R
e
MC 2000
MC 2001
MA 2001
MA 2000
FP 2000
MA 1999
1:1 Line
RegressionMiles City 2000 flux data begins 4/25
ResultsHow Well do 10-day Model Tree Predictions Averaged over
the Growing Season Compare with Flux Tower Observation?
NGP Modeled April - Oct. average Pg (g C m-2 day-1) versus Observed Pg, MAE = 0.15
y = 0.9602x + 0.1677R2 = 0.9049
1
1.5
2
2.5
3
1 1.5 2 2.5 3
Modeled Pg
Tow
er P
g
MC 2000MC 2001MA 2001MA 2000FP 2000MA 19991:1 LineRegression
Miles City 2000 flux data begins 4/25
Twrlocalbxy_gpp_nee01_Re.xls
NGP Modeled April - Oct. average NEE (g C m-2 day-1) versus Observed NEE
y = 1.0274x + 0.0279R2 = 0.9545
-0.4
-0.2
0
0.2
0.4
0.6
0.8
-0.4 -0.2 0 0.2 0.4 0.6 0.8
Modeled NEE
Tow
er N
EE
MC 2000MC 2001MA 2001MA 2000FP 2000MA 19991:1 LineRegression
0
1
2
3
4
5
6
0 1 2 3 4 5 6
Tower Pg (g C m-2 day-1)
Estim
ated
Pg
(g C
m-2
day
-1)
LEMAMC1:1 Line
MAD = 0.32Relative Error = 0.08
0
1
2
3
4
5
6
0 1 2 3 4 5 6
Tower Pg (g C m-2 day-1)
Estim
ated
Pg
(g C
m-2
day
-1)
LEMAMC1:1 Line
MAD = 0.56Relative Error = 0.16
Model Tree (trained from all towers) MODIS GPP (withheld 2001 MAD was 0.48)
Model Tree Predictions Match 2001 Flux Tower ObservationsBetter than MODIS Net Photosynthesis
• Captures General Tend in Tower Pg• Predictions seem site specific
MODIS
Regional Comparison of 2001 MODIS GPP and Model Tree Pg on Rangelands
Pg (g C m-2 day-1)
• General Agreement • MODIS had outlier values > 6 g C m-2 day-1
ModelTree
1998 = -15.5
NEE (g C m2 yr-1)
1999 = 16.82000 = -59.42001 = -6.0Average = -16.0
Ppt_temp_monthly_rangelands.xls0
10203040
5060708090
0 2 4 6 8 10 12
Month
Pre
cipi
tatio
n (m
m)
2001200019991998
Ppt_temp_monthy_rangelands.xls
1998 = -15.5
NEE (g C m2 yr-1)
1999 = 16.82000 = -59.42001 = -6.0Average = -16.0
Why was 2000 NEE so low?
• Low June precipitation
• Low previous fall precipitation
0
10
20
30
40
High 1 High 2 Low 1 Low 2
Perc
ent C
lay
High NEE 1 and 2
Low NEE 1 and 2
Investigating Causes of Sinks and Sources: 1999
-2
-1
0
1
2
3
2/9 3/31 5/20 7/9 8/28 10/17 12/6
Date
NEE
(gC
m-2
day
-1)
High 1High 2Low 1Low 2
0
0.1
0.2
0.3
0.4
0.5
0.6
2/9 3/31 5/20 7/9 8/28 10/17 12/6
Date
ND
VI
High 1High 2Low 1Low 2
STATSGO Range Production Normal Year
Kg ha-1
1998-2001 April-June Precipitation
mm
<= 525
>= 50
Percent clay
STATSGO percent clay
Environmental Drivers of 1998-2001 Source Areas
Average Temperature (Apr. - Oct., 1998-2001)
13.7
15.5
17.4
18.9
(C)
Conclusions
• Model Tree predictions of Pg and NEE agreed well with tower observations
• Heavy reliance of Model Tree predictions on NDVI insures tracking of temporal and spatial dynamics not evident in climatic data sets
• Winter flux estimates need improvement
• N. Great Plains Rangelands from 1998-2001 were a WEAK source of C.