Estimation of Terrestrial Global Gross Primary Production (GPP) with Satellite Data-Driven Models and Eddy Covariance Flux Data

JOANNA JOINER, YASUKO YOSHIDA, YAO ZHANG, GREGORY DUVEILLER, MARTIN JUNG, ALEXEI LYAPUSTIN, YUJIE WANG, COMPTON TUCKER

Joiner et al., Remote Sens. 2018, 10(9), 1346; https://doi.org/10.3390/rs10091346

Outline

u Motivationu Basic Approachu What didn’t work so wellu What did work wellu What are the implications (global GPP)u Summary

Outline

uMotivationu Basic Approachu What didn’t work so wellu What did work wellu What are the implications (global GPP)u Summary

Motivation: What are current approaches to estimate global GPP? strengths/challenges?

u Machine learning typically trains to predict GPP from eddy covariance data (100+ sites, gold standard) with satellite and meteorological data inputs; works well but doesn’t provide a physical basis for understanding; Which are the most important data?

u Light-use Efficiency (LUE) models use satellite and meteorological data: GPP = PAR * fPARchl * LUE,

u fPARchl = fraction of absorbed photosynthetically active radiation)

u how good are the LUE parameterizations (e.g., do they create features that are (not) supported by eddy covariance observations)?

u Terrestrial biology models (TBMs) vary in their performance and need evaluation with data-driven approaches.

u Assimilation and SIF-based estimates (with optional downscaling) are newer approaches, but how good are they? need comparison with other approaches

Carbon fluxes in terrestrial ecosystems and relationship to satellite remote sensing

Net Ecosystem Exchange NEE = GPP + Reco

PARTOA

PARIn

Reflected light -> fPARchl

Solar-induced fluorescence (SIF)

Ecosystem Respiration Reco = Rsoil + Rplants

!

GPP

Motivation: Why do we need another GPP estimate? Questions to consider:

u How well can we do just with satellite reflectances? Angle (BRDF) adjusted?u Use vegetation indices? and from what source, e. g., from composites? u Deriving fAPARchl (Q. Zhang et al.) with MODIS reflectances works well but computationally

intensive to produce. Is there a simple data-driven approach (linear band combination)? u What value can fluorescence (SIF) add? u What are most important LUE drivers? Is there a straight-forward way to compute LUE?u On what time scales can we accurately estimate GPP with satellite data? Monthly? Daily? u What spatial scale is needed to evaluate or train with eddy covariance flux tower data?

Outline

u Motivationu Basic Approach (including input data sets)u What didn’t work so wellu What did work wellu What are the implications (global GPP)u Summary

Basic Light Use Efficiency (LUE) Approach

u LUE concept: GPP = PARIN * fPARchl * LUEp

u Similar for SIF: SIF = PARf * fPARchl,f * LUEf * fesc

u Many studies show GPP-SIF correlation(weekly to monthly time scales, but scaling factor varies)

CERES or assimilation

MODIS or similar

Parameterization(meteorological, satellite data)

Considerations: SIF vs. Reflectances

u Large signal in reflectances (NIR – Red); Easy to measure (low spectral resolution bands), sensitive to clouds

u Small SIF signal requires high spectral resolution -> lower temporal and spatial resolution, but less sensitive to clouds

u Both sensitive to fPARchl, but SIF also sensitive to PARIN.u SIF ~ linear with GPP (weekly to monthly), but scaling factor varies,

complicated by fesc and LUEf

u SIF shows more sensitivity to high yield (C4) crops such as cornu Can we utilize the assets of both?

Modified (simplified) LUE Approach

u 1) GPP = fPARchl * PARIN * LUE , where LUE is function of meteorological parameters.

u 2) GPP =~ fPARchl * PARTOA ( * LUE’), where LUE’ is optional & a function of reflectances

u LUE dependence on PAR is important (diurnal, cloud effects)!u PARTOA ⍺ PARIN * LUE implicitly accounts for LUE variation with PAR to keep GPP nearly

constant over a wide range of PAR (implied by earlier works for croplands by Peng and Gitelson)

u Majority of stress effects are reflected in fPARchl (e.g., leaf curling, loss of chlorophyll) which occurs relatively quickly (within days to a few weeks of stress onset)

u LUE’ can be (optionally) parameterized as a function of fPARchl (or NDVI) to account for additional stress effects (moisture, temperature, radiation, nutrients) – accounts for very small amount of GPP variability

Vegetation indices as proxies for fPARchl: What are the issues?

u NDVI= (RNIR – Rred)/(RNIR + Rred); R: Reflectanceu NDVI /= 0 for background (soils, branches, etc.)u Subtracting an constant offset from NDVI isn’t the right thing to do (isn’t necessary at high NDVI

values). Instead, we subtract and phase out an offset value N0 linearly with NDVI up to NDVI=0.7, i.e., NDVI’=NDVI – f(N0).

u How to determine N0? We tried using SIF (at lower spatial resolution) and GPP at flux sites, but ultimately settled on a constant value of 0.25.

u What about the NDVI “saturation problem” at high NDVI values?u EVI is a similar 3 band index intended to improve upon NDVI (used in VPM)u NIRV=(NDVI – 0.08) * RNIR

u How well does a simple linear combination of band reflectances work better?

FOLIAGEOCO-2,3

680 700 720 740 760 780Wavelength (nm)

0.0

0.5

1.0

1.5

2.0

Scal

ed s

igna

l (Ar

bitra

ry U

nits

)

Reflected RadianceFluorescenceTOA Solar IrradianceAtmospheric Transmittance

O2 A-bandO2 B-band water vapor

Red NIR->

Data sets used for GPP estimationInput Dataset Temporal Resolution Spatial Resolution Use

MCD43D reflectancesSchaaf et al.

daily 0.0083∘×0.0083∘ ~1 km Indices -> fPARCHL

GOME-2 SIF* (downscaled)Duveiller & Cescatti, 2016

monthly 0.05∘×0.05∘ GPP proxy (adjusted for daily PAR)

GOME-2 SIFJoiner et al., 2013, 3016

monthly 0.5∘×0.5∘ Identify high productivity areas

CERES or GEOS-DAS any 0.5∘×0.5∘ PARIN

NIRV x SWTOA any any GPP proxy

NDVI’ x SWTOA any any GPP proxy

FLUXNET 2015 GPP Daily, monthly Site footprint ~1 km Training, evaluation

VPM GPP (LUE model)Y. Zhang et al., 2017

monthly 0.05∘×0.05∘ Baseline evaluation

FLUXCOM GPP(ML) Tramontana et al., 2016

8-day 0.083∘×0.083∘ Baseline evaluation

FLUXNET 2015 Tier 1 eddy covariance sites (free and open use)

u Covers a wide range of ecosystems

u Sparse coverage over some continents and regions (e.g., tropical rain forests, high latitudes)

-100 -80 -60 -40

-40

-20

0DBFMFENFEBFCROOSHSAVCSHGRAWETWSASNO

-160 -120 -80 -40

40

60

0 20 40

40

60

80

-20 0 20 40

-20

0

20

80 120 160

30

60

120 135

-30

-15

Outline

u Motivationu Basic Approachu What didn’t work so wellu What did work wellu What are the implications (global GPP)u Summary

Things we tried that didn’t work so well

u Using MOD09 reflectances (BRDF-adjusted data sets MCD43 NBAR and MAIAC worked better)

u Using MOD15 fPAR (and MOD17 GPP based on it)u Using MOD13 NDVI (max value) composited data setsu Using CERES/MERRA-2 SW as PARIN with constant or simple

LUE (PARTOA works better)u Using NIRV without any account of PARu Accounting for SIF escape (SZA dependence) using simple

MODIS reflectance-based approach

Why does using PARTOA work better?

u Plants are adjusting their efficiency such that GPP stays relatively constant over a wide range of PAR where most of the days fall

(Higher LUE at lower PAR; lower PAR corresponds to cloudy days with more diffuse light)

Symbols for different site-years

15 day period in July (~static vegetation)

Outline

u Motivationu Basic Approachu What didn’t work so wellu What did work wellu What are the implications (global GPP)u Summary

Monthly GPP, 5 km: SIF* versus NDVI- and NIRV-based (evaluation vs independent sites)

y= 1.249*x +-0.394

0 5 10 15 20 25 30SIF*

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

r2 =bias =RMSE =MEF =λ =

0.602 0.5302.666

0.560 0.687

a)

0 5 10 15 20 30 40 50 60 80 100

y= 1.122*x +-0.292

0 5 10 15 20 25 30NDVI’ * SWTOA

0

5

10

15

20

25

30FL

UXNE

T 20

15 G

PP

r2 =bias =RMSE =MEF =λ =

0.762 0.2012.007 0.751 0.845

b)y= 1.166*x + 0.123

0 5 10 15 20 25 30NIRV * SWTOA

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

r2 =bias =RMSE =MEF =λ =

0.707 0.7072.337 0.662 0.782

c)

Zero-offset problem in early version of GOME-2 SIF (now corrected)

Number of points

Mainly sites in US cornbelt

8-day GPP, ~1 km, linear combination of bands versus NDVI-based

Mainly sites in US cornbelt

Number of points

Locations with high (normalized) SIF to NDVI ratios

-150 -120 -90 -60 -30 0 30 60 90 120 150 180-60

-30

0

30

60

• separate fits for

1. these points 2. all others

• Additional parameterization of LUE to account for photoperiod and stress effect on LUE• Tried several approaches including use

of PAR and NDVI• Best results with polynomial in NDVI

Dual fit, variable LUE - ”FluxSat” ( “SatFlux”)

8-day 1 km GPP, dual fit (FluxSat)

Number of points

Monthly, 5 km GPP comparison with VPM

Number of points

8-day, 8 km GPP comparison with FLUXCOM-RS

Training data set matters?! FLUXCOM used an older FluxNet data set.Number of points

How well do the data sets capture interannual variations (anomalies)?

Anomalies normalized by GPP seasonal range (fractional, unitless)

Number of points

How well do data sets capture site-to-site (spatial) variability?

Each point is an annual mean for a single site

DBF+MF (n=4143, r2= 0.830)

0 5 10 15 20 25FluxSat-7

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 0.995*x +-0.386DBF+MF (n=4143, r2= 0.808)

0 5 10 15 20 25NIRV * SWTOA

0

5

10

15

20

25

30FL

UXNE

T 20

15 G

PPy= 0.887*x + 0.197

DBF+MF (n=4143, r2= 0.799)

0 5 10 15 20 25FluxSat-N

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 1.058*x +-0.511

EBF (n=1113, r2= 0.526)

0 5 10 15 20 25FluxSat-7

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 0.714*x + 0.971EBF (n=1113, r2= 0.430)

0 5 10 15 20 25NIRV * SWTOA

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 0.658*x + 1.572EBF (n=1113, r2= 0.581)

0 5 10 15 20 25FluxSat-N

0

5

10

15

20

25

30FL

UXNE

T 20

15 G

PPy= 0.776*x + 0.358

ENF (n=2212, r2= 0.793)

0 5 10 15 20 25FluxSat-7

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 1.244*x +-0.906ENF (n=2212, r2= 0.782)

0 5 10 15 20 25NIRV * SWTOA

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 1.362*x +-0.071ENF (n=2212, r2= 0.816)

0 5 10 15 20 25FluxSat-N

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 1.003*x +-0.302

CSH+OSH (n= 576, r2= 0.642)

0 5 10 15 20 25FluxSat-7

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 0.594*x + 0.211CSH+OSH (n= 576, r2= 0.812)

0 5 10 15 20 25NIRV * SWTOA

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 0.884*x +-0.271CSH+OSH (n= 576, r2= 0.671)

0 5 10 15 20 25FluxSat-N

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 0.573*x + 0.203

WSA+SAV (n= 884, r2= 0.733)

0 5 10 15 20 25FluxSat-7

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 0.984*x + 0.899WSA+SAV (n= 884, r2= 0.800)

0 5 10 15 20 25NIRV * SWTOA

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 1.159*x +-0.225WSA+SAV (n= 884, r2= 0.790)

0 5 10 15 20 25FluxSat-N

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 1.111*x + 0.454

GRA (n=3335, r2= 0.825)

0 5 10 15 20 25FluxSat-7

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 0.930*x + 0.125GRA (n=3335, r2= 0.802)

0 5 10 15 20 25NIRV * SWTOA

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 0.946*x +-0.256GRA (n=3335, r2= 0.818)

0 5 10 15 20 25FluxSat-N

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 1.002*x +-0.074

WET (n= 330, r2= 0.638)

0 5 10 15 20 25FluxSat-7

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 1.230*x + 0.627WET (n= 330, r2= 0.671)

0 5 10 15 20 25NIRV * SWTOA

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 1.239*x + 0.956WET (n= 330, r2= 0.652)

0 5 10 15 20 25FluxSat-N

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 1.311*x + 0.628

CRO (n= 999, r2= 0.818)

0 5 10 15 20 25FluxSat-7

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 1.014*x +-0.039CRO (n= 999, r2= 0.726)

0 5 10 15 20 25NIRV * SWTOA

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 1.230*x +-0.869CRO (n= 999, r2= 0.811)

0 5 10 15 20 25FluxSat-N

0

5

10

15

20

25

30

FLUX

NET

2015

GPP

y= 1.122*x + 0.070

Black line: 1:1 Red line: fit Each row is a different vegetation type

Outline

u Motivationu Basic Approachu What didn’t work so wellu What did work wellu What are the implications (global GPP)u Summary

July 2007 8-day average GPP

July 2007 FLUXCOM-RS FluxSat-7

GPP 8 day average January 2007

FluxSat-7January 2007 FLUXCOM-RS

2007 Annual average GPP

140.8 Pg C / yr125.0 Pg C / yr

Vegetation Photosynthesis Model (VPM) FluxSat-7

Are high FluxSat values in tropics and high latitudes supported by FLUXNET?

0 2 4 6 8 10 FluxSat-7 or VPM GPP (g C m-2 d-1)

0

2

4

6

8

10

GPP

FLUX

NET

(g C

m-2 d-1

)FluxSat-7 WETVPM WETFluxSat ENFVPM ENF

b)

4 6 8 10 12 FluxSat-7 or VPM GPP (g C m-2 d-1)

4

6

8

10

12

GPP

FLUX

NET

(g C

m-2 d-1

)

FluxSat-7

VPM

a)

Tropical site (withheld from training) High latitude sites withheld from training

Annual mean GPP, model comparison (2003-2017)

FluxSat (SatFlux, version 1) is publicly available from the AVDC website:https://avdc.gsfc.nasa.gov

Go to Data and Archive menus

FluxSat

Courtesy Eunjee Lee, Fan-Wei Zeng, Randy Koster GMAO

Summary

u We developed a simplified LUE approach to estimate global GPP with satellite data using assets of MODIS reflectanaces and SIF (FluxSat) trained using latest FLUXNET 2015 data set

u FluxSat performs as well or better than other more complex formulations (as compared with independent FLUXNET data)

u FluxSat estimates 2007 global annual GPP at 140.8 Pg C / yr - generally higher than other satellite-based estimates but comparable to many TBM estimates.

u Still investigating details of multi-year data set and expect improvements in the future.

u Jury still out as to whether or not satellite driven SIF-based estimates can outperform reflectance-based GPP estimates.

u Need more flux towers in under-observed regions such as tropical rain forests

Backups

Correlation of NDVI and root-zone soil moisture (RZM) weekly anomalies (indicated by ‘). Gray areas are where correlations are not statistically significant. Highest correlations in semi-arid regions.

Time lags in days of NDVI with respect to root-zone soil moisture (RZM) weekly anomalies. Positive numbers are where NDVI’ lags RZM’. Typically lags are days to a few weeks.

From Joiner et al., Rem. Sens. Environ., 2018

Correlation

Time Lag

MODISBRDF-

adjustedreflectances(MCD43D)

FLUXCOM-RS or VPM GPP (driven

by MODIS and other data)

GOME-2 SIFSIF*

downscaled with MODIS

FLUXNET 2015 GPP Tier 1 odd numbered sites

(for training)

FLUXNET 2015 GPP Tier 1 even numbered sites(for evaluation)

Top-of-atmosphere

(TOA) shortwave

(SW) irradiance

Our satellite-based GPP estimates

Compare statistically

(END)

Train/calibrate satellite data to estimate

GPPMultiply TOA SW

and FAPAR proxies

Use SIF and NDVI to

identify high productivity

areas (optional)

Calibrate fPARchl

proxies OR compute VIs (NDVI, NIRV )

Add LUE parame-terization(optional)

evaluation

training

LUE