Remote Sensing of Environment 141 (2014) 90–104

Contents lists available at ScienceDirect

Remote Sensing of Environment

j ourna l homepage: www.e lsev ie r .com/ locate / rse

Estimating landscape net ecosystem exchange at high spatial–temporalresolution based on Landsat data, an improved upscaling modelframework, and eddy covariance flux measurements

Dongjie Fu a,b, Baozhang Chen a,b,⁎, Huifang Zhang a,b, JuanWang c, T. Andy Black d, Brian D. Amiro e, Gil Bohrer f,Paul Bolstad g, Richard Coulter h, Abdullah F. Rahman i, Allison Dunn j, J. Harry McCaughey k,Tilden Meyers l, Shashi Verma m

a LREIS, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, 11A, Datun Road, Chaoyang District, Beijing, Chinab University of Chinese Academy of Sciences, Beijing 100049, Chinac Department of Geography and Resource Management, The Chinese University of Hong Kong, Shatin, NT, Hong Kong, Chinad Faculty of Land and Food Systems, University of British Columbia, 2357 Main Hall, Vancouver V6T 1Z4, Canadae Department of Soil Science, University of Manitoba, Winnipeg, Manitoba R3T 2N2, Canadaf Department of Civil, Environmental and Geodetic Engineering, The Ohio State University, 417E Hitchcock Hall, 2070 Neil Avenue, Columbus, OH 43210, USAg Department of Forest Resources, University of Minnesota, 1530 N. Cleveland Ave., St. Paul, MN 55108, USAh Environmental Science Division, Climate Research Section, Argonne National Laboratory, Building 203, 9700 South Cass Avenue, Argonne, IL 60439, USAi Department of Geography, Indiana University, 701 E Kirkwood Avenue, Bloomington, IN 47405, USAj Department of Physical and Earth Sciences, Worcester State University, Worcester, MA 01602-2597, USAk Department of Geography, Queen's University, Mackintosh-Corry Hall, Room E112, Kingston, Ontario, Canada K7L 3 N6l Atmospheric Turbulence and Diffusion Division, NOAA/ARL, P.O. Box 2456 456, South Illinois Avenue, Oak Ridge, TN 37831-2456, USAm School of Natural Resources, University of Nebraska-Lincoln, 807 Hardin Hall, Lincoln, NE 68583-0728, USA

⁎ Corresponding author at: LREIS, Institute of GeoResources Research, Chinese Academy of Sciences,District, Beijing, China. Tel./fax: +86 10 64889574.

E-mail addresses: baozhang.chen@igsnrr.ac.cn, Baozha

0034-4257/$ – see front matter © 2013 Elsevier Inc. All rihttp://dx.doi.org/10.1016/j.rse.2013.10.029

a b s t r a c t

a r t i c l e i n f o

Article history:Received 10 April 2013Received in revised form 24 October 2013Accepted 26 October 2013Available online 22 November 2013

Keywords:Net ecosystem exchangeEddy-covarianceRegression treeImage fusionFootprint climatologyUpscaling

More accurate estimation of the carbon dioxide flux depends on the improved scientific understanding of the ter-restrial carbon cycle. Remote-sensing-based approaches to continental-scale estimation of net ecosystem ex-change (NEE) have been developed but coarse spatial resolution is a source of errors. Here we demonstrate asatellite-based method of estimating NEE using Landsat TM/ETM + data and an upscaling framework. Theupscaling framework contains flux-footprint climatology modeling, modified regression tree (MRT) analysis andimage fusion. By scaling NEEmeasured atflux towers to landscape and regional scales, this satellite-basedmethodcan improve NEE estimation at high spatial-temporal resolution at the landscape scale relative to methods basedon MODIS data with coarser spatial–temporal resolution. This method was applied to sixteen flux sites from theCanadian Carbon Program and AmeriFlux networks located in North America, covering forest, grass, and croplandbiomes. Compared to a similar method using MODIS data, our estimation is more effective for diagnosing land-scape NEE with the same temporal resolution and higher spatial resolution (30 m versus 1 km) (r2 = 0.7548vs. 0.5868, RMSE = 1.3979 vs. 1.7497 g C m−2 day−1, average error = 0.8950 vs. 1.0178 g C m−2 day−1, rela-tive error = 0.47 vs. 0.54 for fused Landsat and MODIS imagery, respectively). We also compared the regionalNEE estimations using Carbon Tracker, our method and eddy-covariance observations. This study demonstratesthat the data-driven satellite-based NEE diagnosed model can be used to upscale eddy-flux observations to land-scape scales with high spatial–temporal resolutions.

© 2013 Elsevier Inc. All rights reserved.

1. Introduction

A number of different methods have been developed to esti-mate net ecosystem carbon exchange (NEE), which can be classi-fied as top-down or bottom-up approaches. Under some pieces of

graphic Sciences and Natural11A, Datun Road, Chaoyang

ng.Chen@ubc.ca (B. Chen).

ghts reserved.

constraining information, such as regional prior flux estimates(e.g. Gurney et al., 2003) or an imposed error covariance structure(e.g. Gourdji, Mueller, Schaefer, & Michalak, 2008; Michalak,Bruhwiler, & Tans, 2004; Mueller, Gourdji, & Michalak, 2008), thetop-down approaches are based on atmospheric CO2 concentrationmeasurements and inverse modeling (Ciais et al., 2010; Deng et al.,2007; Gurney, Baker, Rayner, & Denning, 2002, 2008, Gurney et al.,2002; Hayes et al., 2012; Peters et al., 2010) to estimate the surfaceemissions given observed fields of atmospheric CO2 concentration,wind speed and wind direction. The bottom-up approaches use

Pre-processing

Reflectance—band

1,2,3,4,5,7

LST—band 6

Vegetation index (EVI,

NDWI)

Test data

Training data

Regression tree algorithm

NEE based rules

Image fusion (ESTARFM/STARFM)

Landsat-like reflectance

High temporal resolution

NEE

Pre-processing

High spatial resolution

NEE

RMSE,R^2

Not meet the pre-defined

standard

High spatiotemporal resolution NEE

Meet the pre-defined standard

SAFE-f model

footprint weighted value within 6 km × 6 km centered at flux site

Pre-processing

NEE (flux data)

Land cover type of the

flux site

Landsat TM/ETM+

Flux data & meteorological

data

MODIS LST (MOD11A2)

MODIS reflectance (MOD09A1)

NEE (flux data)

Fig. 1. The flowchart of landscape NEE estimation algorithm.

91D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

observations of the surface properties to resolve CO2 emission rates, in-cluding using emission models based on observations from eddy-covariance (EC) flux-towers, biomass inventories (Peylin et al., 2005;Stinson et al., 2011), terrestrial biosphere models (Hayes et al., 2012;Keenan, Baker, et al., 2012) and remote sensing data products(Churkina, Schimel, Braswell, & Xiao, 2005; Xiao, Zhuang, et al., 2011;Xiao et al., 2008). Progress in estimating carbon fluxes has been achievedat either the large, continental scale(Gurney et al., 2002, 2008) or thelocal, ecosystem scale (typically less than 1–3 km2 for each site) (Chenet al., 2012). However, the landscape-scale (101–102 km2) carbon fluxand especially its spatial–temporal variations remain poorly modeled(Chen, Chen, Mo, Black, & Worthy, 2008; Cook et al., 2004; Piao et al.,2009). Accurate quantification of the NEE dynamics at the landscapeand regional scales is comparatively weak and achieving greater accura-cy and precision of modeling at this level are crucial to improving ourunderstanding of the terrestrial carbon cycle locally and for reducingglobal carbon budget errors (Keenan, Davidson, Moffat, Munger, &Richardson, 2012; Tang et al., 2012; Xiao, Chen, Davis, & Reichstein,2012; Xiao, Davis, Urban, Keller, & Saliendra, 2011; Xiao, Zhuang, et al.,2011; Xiao et al., 2008).

Remote-sensing-based methods have the potential to scale the ECmeasurement of NEE to larger spatial scales. Unlike other bottom-upmethods, remote-sensing-based approaches are not limited by theavailability of the in situ ground measurements. Veroustraete, Patyn,and Myneni (1996) combined normalized difference vegetation index(NDVI) and land surface flux data to estimate NEE using an ecosystemmodel. Maselli, Chiesi, Fibbi, and Moriondo (2008) and Maselli et al.

(2010) combined aircraft EC flux data, remote-sensing-based andprocess-based ecosystem models to estimate NEE at different spatial–temporal scales. Mahadevan et al. (2008) used a satellite-based assimi-lation scheme to estimate NEE based on vegetation indices derived fromModerate Resolution Imaging Spectroradiometer (MODIS) data, climatedata and tower EC flux data. Xiao et al. (2008), Xiao, Zhuang, et al.(2011) estimatedNEE at 1-km and 8-day resolutions over the contermi-nous United States by combiningMODIS imagery and tower ECflux datafrom the AmeriFlux database using the modified regression tree (MRT)method. To our knowledge, there has been no such published study es-timating landscape-scale NEE with high spatial (less than 100 m) andhigh temporal resolutions by making use of available global EC fluxdata and remote sensing imagery.

In this study, we developed an integratedmethod to estimate NEE atlandscape scales with high spatial resolution (30 m) by synthesizing ECflux measurements with remotely-sensed data to account for the landsurface heterogeneity. This approach combines an enhanced spatialand temporal adaptive reflectance fusion model (ESTARFM, Zhu, Chen,Gao, Chen, & Masek, 2010), a Simple Analytical Footprint model onEulerian coordinates (SAFE-f, Chen, Black, Coops, Hilker, et al., 2009;Chen et al., 2010, 2012) and the MRT method.

This approach employs these assumptions: i) only the target land-cover type observed by the flux tower (of which the flux footprint is typ-ically less than 1–3 km2, Chen et al., 2010) is taken as the contribution ofobserved carbon flux and selected for upscaling using a footprint model,ii) the variation of phenology and the difference between the nearestavailable Landsat surface reflectance data and the corresponding MODIS

Fig. 2. Locations of flux sites and the spatial distributions of land cover (MCD12Q1, year: 2001) in North America. For description of flux sites, see Table 1.

92 D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

8-day surface reflectance products during a short time interval (i.e.8 days) are small enough to be ignored, and iii) the land surface temper-ature (LST) data derived from Landsat (around local noon time) can rep-resent the average temperature status of land surface over 8-day period.

2. Data and methods

Fig. 1 presents the flowchart of the landscape NEE estimation algo-rithmwith high spatial–temporal resolution. First, the probability distri-bution functions of the EC flux footprint for the corresponding Landsator fused Landsat-like imagery periods (an 8-day interval, 2000–2006)were calculated using the footprint model—SAFE-f (Chen, Black,Coops, Hilker, et al., 2009), and then the Landsat data were weightedby a flux footprint probability density function (PDF) for the area sur-rounding the tower. The fused Landsat-like imagery was produced by

Table 1Descriptions of flux sites in this study. Site ID and name: First two letters indicate the province (United States.

Land covertype

Site ID Site name Longitude Latitude Fr

Needleleafforest

CA-Obs SK-Southern Old Black Spruce −105.1178 53.9872 2CA-Ojp SK-Old Jack Pine −104.6920 53.9163 2CA-Ca1 BC-Campbell River 1949 Douglas-fir −125.3351 49.8688 2CA-Man MB-Northern Old Black Spruce −98.4808 55.8796 2

Broadleafforest

CA-Oas SK-Old Aspen −106.1978 53.6289 2US-MMS Morgan Monroe State Forest −86.4131 39.3232 2US-UMB Univ. of Mich. Biological Station −84.7138 45.5598 2US-WCr Willow Creek −90.0799 45.8059 2

Cropland US-Bo1 Bondville −88.2904 40.0062 2US-Ne1 Mead - irrigated continuous maize site −96.4766 41.1651 2US-Ne2 Mead - irrigated maize-soybean

rotation site−96.4701 41.1649 2

US-Ne3 Mead-rainfed maize-soybeanrotation site

−96.4396 41.1797 2

Grass US-FPe Fort Peck −105.1019 48.3077 2US-Var Vaira Ranch −120.9507 38.4067 2US-Aud Audubon Research Ranch −110.5092 31.5907 2US-Wlr Walnut River Watershed −96.855 37.5208 2

an image fusion model which will be used at the third step. The imagefusion model, described in details at Section 2.3.3, was used to produceLandsat-like imagery after blending Landsat TM/ETM+ andMODIS im-agery. Second, the regression-based NEE diagnosed model, which com-bined with several linear regression models and subjected to certainlimits (see Section 2.3.1), was produced and then calibrated by MRT.The data used for the diagnosed model were the weighted reflectance,derived vegetation index, LST of Landsat imagery and the correspondingEC flux data. The period for training set and test set of MRT were 2000–2004 and 2005–2006, respectively. Third, the landscape NEE at 8-dayintervals was estimated using the regression-based NEE diagnosedmodel based on the combination of original and fused Landsat surfacereflectance data using an image fusion model (ESTARFM), vegetationindex derived from original and fused Landsat reflectance, Landsat LSTand MODIS LST data. The MODIS LST data were combined with fused

MB—Manitoba, SK— Saskatchewan, BC— British Columbia) of Canada (CA), US indicates

lux dataange

Path Row Reference

000–2006 37 22 Barr, Morgenstern, Black, McCaughey, and Nesic (2006)000–2006 37 22 Barr et al. (2006)000–2006 49 25 Chen, Black, Coops, Hilker, et al. (2009)000–2006 34 21 Dunn, Barford, Wofsy, Goulden, and Daube (2007), Turner

et al. (2003)000–2006 38 23 Barr et al. (2006)000–2006 21 33 Dragoni, Schmid, Grimmond, and Loescher (2007)000–2006 21/22 28 Nave et al. (2011)000–2006 25 28 Cook et al. (2004)000–2006 22/23 32 Hollinger, Bernacchi, and Meyers (2005)001–2006 28 31 Suyker and Verma (2008)001–2006 28 31 Suyker and Verma (2008)

001–2006 28 31 Suyker and Verma (2008)

000–2006 35/36 26 Wilson and Meyers (2007)001–2006 43/44 33 Ryu, Baldocchi, Ma, and Hehn (2008)002–2006 35/36 38 Wilson and Meyers (2007)002–2004 27/28 34 LeMone et al. (2002)

2 1 0 1 22

1

0

1

2(a1) CAObs

2 1 0 1 22

1

0

1

2 (a2) CAOjp

2 1 0 1 22

1

0

1

2 (a3) CACa1

2 1 0 1 22

1

0

1

2(a4) CAMan

2 1 0 1 22

1

0

1

2 (b1) CAOas

2 1 0 1 22

1

0

1

2 (b2) USMMS

2 1 0 1 22

1

0

1

2(b3) USUMB

2 1 0 1 22

1

0

1

2(b4) USWCr

2 1 0 1 22

1

0

1

2 (c1) USBo1

2 1 0 1 22

1

0

1

2 (c2) USNe1

2 1 0 1 22

1

0

1

2(c3) USNe2

2 1 0 1 22

1

0

1

2(c4) USNe3

2 1 0 1 22

1

0

1

2 (d1) USFPe

2 1 0 1 22

1

0

1

2(d2) USVar

2 1 0 1 22

1

0

1

2(d3) USAud

2 1 0 1 22

1

0

1

2(d4) USWlr

Distance from tower W E (km)

Dis

tanc

e fr

om to

wer

S N

(km

)

Shadow

Water

Exposed land

Shrub tall

Shrub low

Wetland-tree

Wetland-shrub

Wetland-herb

Herb

Coniferous Dense

Coniferous Open

Coniferous Sparse

Broadleaf Dense

Broadleaf Open

Broadleaf Sparse

Mixedwood Dense

Mixedwood Open

Open Water

Developed, Open Space

Developed, Low Intensity

Developed, Medium Intensity

Developed, High Intensity

Barren Land (Rock/Sand/Clay)

Deciduous Forest

Evergreen Forest

Mixed Forest

Shrub/Scrub

Grassland/Herbaceous

Pasture/Hay

Cultivated Crops

Woody Wetlands

Emergent Herbaceous Wetlands

(NF-1) CA-Obs (NF-2) CA-Ojp (NF-3) CA-Ca1 (NF-4) CA-Man

(BF-1) CA-Oas (BF-2) US-MMS (BF-3) US-UMB (BF-4) US-WCr

(CL-1) US-Bo1 (CL-2) US-Ne1 (CL-3) US-Ne2 (CL-4) US-Ne3

(GR-4) US-Wlr(GR-3) US-Aud(GR-2) US-Var(GR-1) US-FPe

Land cover types for CA sites Land cover types for US sites

Fig. 3. Land covermaps (4.2 km × 4.2 km, 30-mresolution) of sixteenflux sites centered at each tower location. The contours from inner to outer are the 50%, 80%, 95% and99% contours of8-day cumulative footprint climatology PDFs, respectively. Abbreviations for each site are given in Table 1. The footprint data sequence of 8-day for flux sites are 23 (Jun. 26) 2001(US-Ne1,US-Ne2, US-Ne3), 2004 (US-Wlr), 2005 (US-Var) and in 2006 (other sites). The number-23 (Jun. 26) is 23rd 8-day of year (one year contains 46 8-day). NF: needleleaf forest, BF: broadleafforest, GR: grassland, and CL: cropland.

93D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

Landsat reflectance data. Combined with previous three steps, it calledLandsat-based diagnosed model compared to MODIS-based diagnosedmodel which using surface reflectance, vegetation index and LST de-rived fromMODIS data. These are both bottom-upmodels while CarbonTracker (CT) (Peters et al., 2007; Zhang, Chen, & Peters, 2013) discussedbelow is a top-down model. Finally, the diagnosed NEE was evaluatedusing tower EC flux data and compared to CT.

2.1. Study area and data

2.1.1. Study areaWe selected a west–east continental region of North America con-

taining sixteen representative flux towers, where include the mostparts of Canada and the contiguous US (Fig. 2).Wewere able to take ad-vantage of the established flux stations in the AmeriFlux and the

Fluxnet-Canada/Canadian Carbon Program networks. This region con-tains four plant functional types (Table 1): needleleaf forest (4 sites),broadleaf forest (4 sites), grassland (4 sites) and cropland (4 sites),which were reclassified from ten broad classes based on the Universityof Maryland classification system (Friedl et al., 2002). Specifically,needleleaf forest type includes the evergreen needleleaf forest (ENF)and deciduous needleleaf forest (DNF); the broadleaf forest type in-cludes evergreen broadleaf forest (EBF) and deciduous broadleaf forest(DBF); the grassland type includes closed shrublands (CSH), openshrublands (OSH), wood savannas (WSA), savannas (SAV) and grass-land (GRA); the cropland type includes cropland (CRO).

2.1.2. DataThe datasets used in this study include the following four types for

the period 2000–2006: i) Landsat TM/ETM+ reflectance and LST data,

Table 2The sub-models for the regression-based NEE simulation (g C m−2 day−1). The abbreviations—NF, BF, GR, CL denote needleleaf forest, broadleaf forest, grass, and cropland, respectively.The coefficient definitions can be found in Eq. (1).

ID Condition NEE simulation (g C m−2 day−1)

1 R1 ≤ 0.047516, LST N 23.799, LST ≤ 26.476, EVI N 0.33611, land cover in (CL) NEE = −2.3967299 − 385.1 R3 + 236.9 R2 + 145.9 R1 − 44.7 EVI + 39.6 R5 + 33.5R4 − 33.9 R7

2 R1 ≤ 0.047516, LST N 26.476, land cover in (CL) NEE = 43.1281896 − 1.792 LST + 34.6 R2 − 26.7 R3 + 3.4 R1 − 1 EVI − 1.4 R4 + 0.9R5 − 0.8 R7

3 R7 N 0.11248, LST ≤ 23.799, EVI N 0.33611, land cover in (CL) NEE = 15.7090673 − 692.9 R3 + 354.8 R2 + 262.4 R1 − 89.3 EVI + 70 R5 + 73.8R4 − 59.8 R7 − 5.1 NDWI

4 R5 N 0.19474, R7 ≤ 0.11248, LST ≤ 23.799, EVI N 0.33611 NEE = −9.8395327 + 147.9 R7 − 69 R3 + 68.5 R2 − 16 EVI + 6.6 R5 − 3.7 R4 + 0.8 NDWI5 R7 N 0.052542, EVI N 0.33611, NDWI N 0.33023, NDWI ≤ 0.41188, land cover

in (BF, GR, NF)NEE = 8.7090229 − 85.8 R7 − 23.7 NDWI + 42.4 R2

6 R1 N 0.047516, LST N 23.799, EVI N 0.33611, land cover in (CL) NEE = 23.5193427 − 819.1 R3 + 571.9 R2 − 86.1 EVI + 166.5 R1 + 44.4 R5 + 46.8R4 − 37.9 R7

7 EVI N 0.33611, NDWI N 0.41188, land cover in (BF, GR, NF) NEE = 1.8582904 − 44.6 R5

8 R5 ≤ 0.19474, LST ≤ 23.799, EVI N 0.33611, land cover in (CL) NEE = −13.0236512 + 211.3 R7 − 16.8 EVI − 29.1 R3 + 13.2 R2 + 12.5 R1 + 3.3 R5 + 3.5R4

9 EVI N 0.33611, land cover in (GR) NEE = −4.9377598 + 83.3 R5 − 74.5 R7 + 12 NDWI + 30.9 R2 − 23.3 R4 − 8.7 EVI10 R7 ≤ 0.052542, EVI N 0.33611, NDWI ≤ 0.41188 NEE = −7.3993489 + 50.5 R5 + 51.4 R7 + 8 NDWI − 19 R4 + 13.3 R2 − 0.034 LST − 1.9

EVI11 LST N 6.7781, EVI ≤ 0.33611, NDWI N 0.0003819, land cover in (GR, NF) NEE = 0.7230217 + 110.6 R3 − 102.8 R2 − 18.6 R7 − 3.6 R4 + 2.4 R5 + 0.5 NDWI + 0.5

EVI + 0.4 R1

12 R7 N 0.052542, EVI N 0.33611, NDWI ≤ 0.33023, land cover in (BF, NF) NEE = 0.4387566 + 27.3 R5 − 27.4 R7 − 8.6 R4 + 10.5 R2 − 2.2 NDWI − 3 EVI + 4.5R3 − 1.5 R1

13 LST ≤ 6.7781, EVI ≤ 0.33611, NDWI N 0.0003819, land cover in (GR, NF) NEE = −0.2237048 + 0.9 NDWI + 2.7 R3 − 2.2 R4 + 1.5 R5 − 0.8 R7 + 0.3 EVI − 0.8 R2

14 LST N 19.474, EVI ≤ 0.33611, land cover in (GR, NF) NEE = −3.7083819 + 101.9 R3 + 28.7 EVI − 61.3 R4 − 40.7 R2 + 16.8 R5 − 13.4R7 + 0.069 LST

15 LST ≤ 19.474, NDWI ≤ 0.0003819, land cover in (GR, NF) NEE = 0.22532516 R1 N 0.053207, EVI ≤ 0.24045, land cover in (CL) NEE = −0.350819 + 6.3 EVI − 13.5 R2 + 10.2 R1 − 8.6 R7 + 6.3 R5 + 4.3 R3 + 0.015 LST17 R1 N 0.053207, EVI N 0.24045, EVI ≤ 0.33611, land cover in (CL) NEE = 12.9306716 − 37.6 EVI + 56.9 R1 − 51.3 R3 − 1 R4 + 0.7 R5 − 0.6 R7

18 EVI ≤ 0.33611, land cover in (BF) NEE = 0.6354299 − 32.5 R7 + 27.5 R5 + 8.4 R3 − 5.7 R4 − 7 R2 + 5.2 R1 + 0.008 LST19 R1 ≤ 0.053207, EVI ≤ 0.33611, land cover in (CL) NEE = –8.2499909 + 223.1 R1 + 3.1 R3 − 2.1 R4 + 1.4 R5 − 1.2 R7 + 0.4 EVI + 0.2

NDWI − 0.6 R2

Table 3The statistical comparison of the diagnosed overall NEE (g C m−2 day−1) using differentdata and models with the sixteen EC towers measurements.

Data used r2 RMSEb p-Value εα b εr Slope Interceptb

Landsat 0.6445 1.6681 b0.0001 1.0178 0.54 0.7189 −0.1446MODIS 0.5868 1.7497 b0.0001 0.9726 0.54 0.6984 −0.0816Fused Landsata 0.7548 1.3979 b0.0001 0.8950 0.47 0.8291 −0.0318

a Fused Landsat data contain original Landsat data and Landsat-like data produced bythe image fusion method.

b Units: g C m−2 day−1.

94 D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

ii) MODIS reflectance, LST and land cover type data, iii) NEE data obtain-ed fromECflux towers, and iv)meteorological datameasured at thefluxtowers for the SAFE-f model inputs.

Landsat TM/ETM+ level 1 products around the EC flux towers(6 km × 6 km)were acquired from the United States Geological Survey(USGS) EarthExplorer (http://edcsns17.cr.usgs.gov/NewEarthExplorer/)for the period 2000–2006. The path and row of Landsat data for thecorresponding flux sites are shown in Table 1. Co-registration andorthorectificationwere processed for Landsat TM/ETM+using the auto-mated registration and orthorectification package (AROP) (Gao, Masek,&Wolfe, 2009). Subsequently, the digital number (DN) values of LandsatTM/ETM+were calibrated and atmospherically corrected to surface re-flectance using the Landsat Ecosystem Disturbance Adaptive ProcessingSystem (LEDAPS) (Masek et al., 2006). For Landsat thermal band, DNvalues are converted to radiance values using the bias and gain valuesspecific to the individual scene, then convert radiance data to LST apply-ing the inverse of the Planck function (Sobrino, Jiménez-Muñoz, &Paolini, 2004). The cloud and snow assessmentmethod used for LandsatTM/ETM+data correctionwas the Automated Cloud-Cover Assessment(ACCA) algorithm (Irish, 2000; Irish, Barker, Goward, & Arvidson, 2006)when using LEDAPS.

For MODIS data, MOD09A1 (8-day reflectance, 500 m, Collection 5),MOD11A2 (8-day land surface temperature, 1000 m, Collection 5), andMCD12Q1 (annual land cover type, 500 m, Collection 51) around theEC flux towers (6 km × 6 km) were obtained from the National Aero-nautics and Space Administration (NASA) Reverb portal (http://reverb.echo.nasa.gov/reverb/). The MODIS data were re-projected, resampledand clipped to the same resolution and spatial extent as Landsat projec-tion (6 km × 6 km)using theMODIS reprojection tools (available at URLhttps://lpdaac.usgs.gov/tools/modis_reprojection_tool). The invalid datawere eliminated using the quality assurance (QA) flags included in theproduct. For the MOD09A1 data, the QA layer named sur_refl_qc_500mwas used, and the QA flags in binary for each reflectance band variedfrom 0000 to 1111. Within each reflectance band, only the pixels withQA flag 0000 (highest quality) were selected. For the MOD11A2 data,

the QA layers named QC_Day and QC_Night were used, and the QAflags in binary used for LST were mandatory QA flags in the study. Thevalue of mandatory QA flags varied from 00 to 11. For MCD12Q1 data,the QA layer-Land_Cover_Type_QC was used, and the QA flags in binaryused for LST were mandatory QA flags in the study. The value of manda-tory QA flags varied from 00 to 11. For land cover data, daytime andnighttime LST, only the pixels with mandatory QA flag 00 (highest qual-ity) were selected. If no high quality values existedwithin 6 km × 6 km,the period was treated as missing.

The flux tower data were obtained from the Canadian Carbon Pro-gram (CCP) (http://www.fluxnet-canada.ca) and the AmeriFlux(http://public.ornl.gov/ameriflux/dataproducts.shtml) networks. Forthe NEE data used in the study, a negative value denotes carbon uptakeby the ecosystem, and a positive value denotes a carbon source. For theAmeriFlux data, there are two types of NEE data sets: standardized(NEE_st) and original (NEE_or) NEE from level-4 products. The level-4products contain friction-velocity-filtered, gap-filled (Artificial NeuralNetwork (ANN) and Marginal Distribution Sampling (MDS) methods)records, complete with the calculated gross primary productivity andtotal ecosystem respiration terms. Both NEE data sets were gap-filledusing the ANN method (Papale & Valentini, 2003) and/or the MDSmethod (Reichstein et al., 2005). The NEE data of the AmeriFlux siteswith ANN filling were used because ANN method is slightly superiorto MDS method according to Moffat et al. (2007) and Xiao et al.

-15 -10 -5 0

-18

-16

-14

-12

-10

-8

-6

-4

-2

0

2

Observed NEE (g C m-2day-1) Observed NEE (g C m-2day-1)

1:1Needleleaf forestBroadleaf forestGrassCropland

a) Landsat

Dia

gn

ose

d N

EE

(g

C m

−2d

ay−1

)

Dia

gn

ose

d N

EE

(g

C m

−2d

ay−1

)

-15 -10 -5 0 5

-15

-10

-5

0

51:1Needleleaf forest

Broadleaf forestGrassCropland

b) MODIS

Fig. 4. Comparison of observed and diagnosed 8-day NEE using Landsat (a) and MODIS (b) data. The dash line is the 1:1 line.

95D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

(2008). For the CCP NEE data, data gaps were filled following Barr et al.(2004, 2007). The canopy height (hc), EC sensor height (hm), friction ve-locity (u*), friction velocity threshold for EC flux calculation (u*th), hori-zontal wind speed (u), wind direction (WD), standard deviation ofhorizontal wind speed (σv), sensible heat flux (H) and latent heat flux(LE) measured half-hourly at EC sensor height required by the SAFE-ffootprint model were obtained from the same database as NEE. Thevalue of u*th was obtained by visually examining the plot of nighttimeCO2 fluxes versus u*, and locating where the flux begins to level off asu* increases (Chen et al., 2012; Gu et al., 2005). In this approach, missinghalf-hourlymeteorological datawere filled using interpolation from ad-jacent levels in the measurement profile (e.g. air temperature profile)and the average diurnal method for the days before and after the gapwith a 15-day moving window in 5-day increments (Chen, Black,Coops, Krishnan, et al., 2009). The roughness length (z0)was roughly es-timated as 10% of hc (Raupach, 1994).

2.2. Remote-sensing-related parameters and variables

The value of NEE is influenced bymany factors including physical,atmospheric, hydrologic, physiological and edaphic variables (Xiaoet al., 2008). We assumed these factors could be assessed by the in-formation derived from satellite remote-sensing data: land surface

-1.8 -1.6 -1.4 -1.2 -1 -0.8 -0.6 -0.4 -0.2 0-1.8

-1.6

-1.4

-1.2

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

NF

BF

GR

CL

1:1a) Landsat

Observed NEE (g C m-2day-1)

Dia

gn

ose

d N

EE

(g

C m

−2d

ay−1

)

Fig. 5. Comparison of observed and diagnosedmean NEE for the four land cover groups in the sGR: grassland, and CL: cropland. Error bars of land cover are the standard deviation of the obse

reflectance, enhanced vegetation index (EVI), normalized differencewater index (NDWI) and LST. All these variables were derived fromLandsat TM/ETM+ imagery.

The reflectance of a particular land cover type depends on wave-length coverage, soil moisture, sun-object-sensor geometry, biophysicaland biochemical properties (stand age, pigment composition, biomassetc.) (Penuelas, Gamon, Griffin, & Field, 1993; Penuelas, Garbulsky, &Filella, 2011; Penuelas, Garbulsky, Gamon, Inoue, & Filella, 2011;Ranson, Daughtry, Biehl, & Bauer, 1985). The land surface reflectancevaries with view zenith, solar zenith and relative azimuth angle whichlead to brighter or darker spectral observations as a result of the interac-tion of solar irradiance with a given surface and sun-sensor geometry(Roujean, Leroy, & Deschamps, 1992). NDVI is a widely used vegetationindexwhich is often applied in production efficiency models because ofits close correlation with the fraction of photosynthetically active radia-tion (fPAR) absorbed by vegetation and the fractional vegetation coverand other vegetation-related variables (Asrar, Fuchs, Kanemasu, &Hatfield, 1984; Choudhury, 1987; Myneni, Hall, Sellers, & Marshak,1995; Sellers, Berry, Collatz, Field, & Hall, 1992; Zhang et al., 2007).However, NDVI has several limitations including its sensitivity to soilbackground conditions, atmospheric conditions, and saturation in amultilayered, closed canopy (Huete et al., 2002; Xiao et al., 2004,2008). We also used EVI, an improved vegetation index (Huete, Liu,

-1.6 -1.4 -1.2 -1 -0.8 -0.6 -0.4 -0.2 0 0.2-1.6

-1.4

-1.2

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

NF

BF

GR

CL

1:1b) MODIS

Observed NEE (g C m-2day-1)

Dia

gn

ose

d N

EE

(g

C m

−2d

ay−1

)

tudy area using Landsat (a) andMODIS (b) data. NF: needleleaf forest, BF: broadleaf forest,rved and diagnosed 8-day mean NEE.

Table 4The statistical parameters in the comparison of the diagnosed NEE (g C m−2 day−1) using Landsat andMODIS datasets withmeasurements at each of the fifteen EC towers (2005–2006).

Land cover type Site ID Slope Intercepta r2 RMSEa p-Value

Landsat MODIS Landsat MODIS Landsat MODIS Landsat MODIS Landsat MODIS

Needleleaf forest CA-Obs 1.08 1.30 −0.03 −0.05 0.70 0.68 0.51 0.57 b0.0001 b0.0001CA-Ojp 0.688 0.70 −0.03 −0.02 0.49 0.57 0.42 0.39 b0.0001 b0.0001CA-Ca1 1.06 0.81 0.45 −0.37 0.17 0.24 1.50 1.43 b0.0001 b0.0001CA-Man 0.96 0.87 0.06 0.12 0.61 0.56 0.43 0.48 b0.0001 b0.0001

Broadleaf forest CA-Oas 1.28 1.31 −0.40 −0.34 0.87 0.76 1.26 1.54 b0.0001 b0.0001US-MMS 1.02 0.97 −0.29 −0.22 0.89 0.87 1.05 1.11 b0.0001 b0.0001US-UMB 0.92 0.96 0.05 0.0032 0.81 0.80 1.17 1.16 b0.0001 b0.0001US-WCr 1.19 1.03 0.10 −0.61 0.90 0.72 1.30 2.04 b0.0001 b0.0001

Cropland US-Bo1 0.73 0.54 −0.03 −0.31 0.66 0.46 2.14 3.05 b0.0001 b0.0001US-Ne1 0.78 0.80 0.88 0.61 0.81 0.60 2.34 2.74 b0.0001 b0.0001US-Ne2 0.78 0.81 0.51 0.07 0.91 0.55 1.95 3.01 b0.0001 b0.0001US-Ne3 0.73 1.47 0.26 −0.64 0.89 0.84 1.96 2.02 b0.0001 b0.0001

Grass US-FPe 0.43 0.13 −0.06 0.08 0.06 0.002 1.16 1.13 0.018 0.705US-Var 1.15 1.08 0.49 0.06 0.68 0.72 1.00 0.83 b0.0001 b0.0001US-Aud 1.13 0.38 −0.98 −0.57 0.58 0.14 1.25 1.79 b0.0001 0.0003

a Units: g C m−2 day−1.

96 D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

Batchily, & vanLeeuwen, 1997), to overcome the limitations of NDVI.Compared to NDVI which is chlorophyll responsive, EVI is more sensi-tive to canopy structural variations, such as leaf area index (LAI), canopyarchitecture, canopy type and plant physiognomy (Huete et al., 2002).NDWI is defined as the normalized difference between near infrared(NIR) and shortwave infrared (SWIR) spectral bands (Gao, 1996)which is sensitive to vegetation water content and soil moisture. ForLandsat TM/ETM+, corresponding NIR and SWIR bands are band 4(0.78–0.90 μm) and 5 (1.55–1.75 μm), respectively (Jackson et al.,2004). It would be true that ecosystem respiration (Re) likely has strongrelationship to LST, because both autotrophic respiration (Ra) and het-erotrophic respiration (Rh) are known to be significantly controlled byair and near-surface soil temperatures (Xiao et al., 2008; Yuan et al.,2011). The satellite-derived LST has been found to be strongly relatedto Re, especially for densely forested flux sites (Rahman, Sims,Cordova, & El-Masri, 2005).

2.3. Simulation approaches

As mentioned above, we used a method for scaling EC-derived NEEto landscape scales at high spatial (30 m) and high temporal (8-day in-tervals) resolutions by combining a regressionmodel, a footprint clima-tology model, and image-fusion model. This regression-based approachwas used to estimate NEE based on trained rules. The trained rules were

-2.5 -2 -1.5 -1 -0.5 0 0.5 1-2.5

-2

-1.5

-1

-0.5

0

0.5

1

CA-Ca1

CA-Man

CA-Oas CA-Ojp

CA-Obs

US-Aud

US-Bo1

US-FPe

US-MMSUS-Ne1

US-Ne2

US-Ne3US-UMB

US-Var

Observed NEE (g C m-2day-1)

1:1a) Landsat

Dia

gn

ose

d N

EE

(g

C m

−2d

ay−1

)

Fig. 6. Comparison of observed and diagnosedmean NEE for flux sites in the study area using Laobserved and diagnosed 8-day mean NEE. Abbreviations for each site can be obtained in Table

mathematical conditional expressions for NEE estimation produced bycomparing a training data set to observe NEE. The input variablescontained in training set and test set for the diagnosed model includethe surface visible/NIR/SWIR reflectance (6 bands), LST, EVI and NDWIderived from the Landsat imagery, which were weighted by a footprintPDF estimated using the SAFE-f footprint model. A Landsat-like reflec-tance dataset was produced using the image-fusion method (seeSection 2.3.3) to fill any gaps in Landsat data that were contaminatedby clouds or not available.

2.3.1. The regression-based diagnosed modelThe regression tree is a method which can account for the nonlinear

relationship between diagnosed and target variables by producing aregression-based diagnosed model containing more than one rule. Arule is amultivariate linear regression sub-model subject to a set of con-ditions. Both continuous and discrete variables can be set as the inputdata for a regression tree (Xiao et al., 2008; Yang, Huang, Homer,Wylie, & Coan, 2003; Yang, Xian, Klaver, & Deal, 2003). Themodified re-gression tree approach used in this study is the commercial softwarepackage called Cubist (RuleQuest, 2012), linked to R, an open-sourcestatistical computing software environment. Cubist is one type ofregression-based method which was developed by Quinlan (1992).The regression-based diagnosed model for NEE (the target variable) isproduced by Cubist based on EC flux and satellite data as inputs.

Observed NEE (g C m-2day-1)-2 -1.5 -1 -0.5 0 0.5 1

-2

-1.5

-1

-0.5

0

0.5

1

CA-Ca1

CA-ManCA-Oas CA-Ojp

CA-Obs

US-Aud

US-Bo1

US-FPe

US-MMS

US-Ne1US-Ne2

US-Ne3

US-UMB

US-Var

1:1b) MODIS

Dia

gn

ose

d N

EE

(g

C m

−2d

ay−1

)

ndsat (a) andMODIS (b) data. The error bars for each site are the standard deviation of the1. The temporal coverage of the flux data for the plot is provided in Fig. 7.

97D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

The inputs were weighted using the EC flux footprint PDF. Theinput variables were divided into two parts, a training set (theyears 2000–2004, 16 flux sites, 1083 data points) and a test set(the years 2005–2006, 15 flux sites, 363 data points).

2.3.2. Footprint weighted NEEMeasured EC towerfluxes (i.e. NEE) representing the integrated flux

within the tower footprint area (typically 1–3 km2) cannot simply becompared with the remotely sensed NEE derived from the Landsat(30-m) or MODIS (1-km) data because of mismatching at the spatialscale. We used a footprint weighting method to match the Landsatpixel-scale and the EC measurement-scale and hence to optimize the

Fig. 7.Observed and diagnosed 8-dayNEE (g C m−2 day−1) for eachflux site over the period 20represent the observed values, the black solid linewith red symbols (circle and triangle) represbols represents the diagnosed NEE using MODIS data. Abbreviations for each site can be obtacropland.

regression-based NEE diagnosed model. The landscape-scale NEE(FNEE,ϕ) was up-scaled from the spatially distributed 30-m NEE field(FNEE,rs) using

FNEE;ϕ ¼ ∬ΩΠ

FNEE;rs x; yð Þφpure x; yð Þdxdy

FNEE;rs ¼ a1R1 þ a2R2 þ a3R3 þ a4R4 þ a5R5 þ a6R7 þ a7LST þ a8EVI þ a9NDWI

8<:

ð1Þ

where φpure is the pure flux footprint (Chen, Black, Coops, Hilker, et al.,2009),ΩΠ is the upwind footprint source area for the 95% accumulativefootprint, a1, ⋯, a9 are parameters obtained using the regression-based

05–2006using Landsat (contained Landsat-like) andMODIS data. The black circle symbolsents the diagnosed NEE using Landsat data, and the dashed linewith blue pentagram sym-ined in Table 1. (a) NF: needleleaf forest, BF: broadleaf forest, (b) GR: grassland, and CL:

Fig. 7 (continued).

98 D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

NEE diagnosed model, R1, R2, R3, R4, R5, R7 are the reflectance values foreach respective Landsat band, LST is the land surface temperature de-rived from Landsat imagery, and EVI, NDWI are two vegetation indicesderived from the Landsat reflectance data. The FNEE,rs was estimatedfrom Landsat data using the regression-based NEE diagnosed modelwith 30-m resolution. The scale of EC-derived NEE can be matchedusing the footprint-integrated FNEE,ϕ. The land cover maps (Fig. 3) forCanadian sites were acquired from the Earth Observation for Sustain-able Development (EOSD) land cover product of the Canadian ForestService and the Canadian Space Agency (Wulder et al., 2003, 2008),and for the US sites data they were obtained from the National LandCover Dataset (NLCD, 2006) from the USGS EarthExplorer (Fry et al.,2011; Wickham et al., 2013; Xian, Homer, & Fry, 2009).

2.3.3. Satellite image fusionWeproduced a Landsat-like reflectance dataset, which involved two

models: the Spatial and Temporal Adaptive Reflectance Fusion Model(STARFM) (Gao, Masek, Schwaller, & Hall, 2006) and the EnhancedSTARFM (ESTARFM) (Zhu et al., 2010). The STARFM blends LandsatTM/ETM+ images and MODIS images to simulate daily surface reflec-tance at Landsat spatial resolution and MODIS temporal frequency(Gao et al., 2006). Compared to the traditional image fusion methodswhich produce newmultispectral high-resolution imageswith differentspatial and spectral characteristics, such as principle component analy-sis (PCA) (Metwalli, Nasr, Allah, El-Rabaie, & Abd El-Samie, 2010; Naidu& Raol, 2008), intensity-hue-saturation (IHS) (Choi, 2006; Tu, Su, Shyu,& Huang, 2001), Brovey transform (Zhang, 2004) and wavelet

-3

-2.5

-2

-1.5

-1

-0.5

0

0.5

1

An

nu

al a

vera

ged

NE

E (

g C

m-2

day

-1)

CA-Obs

CA-Ojp

CA-Ca1

CA-Man

CA-Oas

US-MMS

US-UMB

US-WCr

US-FPe

US-Var

US-Aud

US-Bo1

US-Ne1

US-Ne2

US-Ne3

Needleaf forest Broadleaf forest Grass Cropland

Flux site

LandsatEC observedMODIS

Fig. 8. Annual averaged NEE (year: 2005) for each flux site according to two methods (Landsat- and MODIS-based diagnosed NEE) and EC observed measurements.

99D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

transform (Li, Manjunath, & Mitra, 1995; Nunez et al., 1999), STARFMhas several advantages: for example, the outputs of traditional image fu-sion are not calibrated to spectral radiance or reflectance (Gao et al.,2006), and may therefore not capture the vegetation changes due tophenological cycles (Gao et al., 2006; Zhu et al., 2010).

The ESTARFM is an improved version of STARFM, which can obtainbetter simulated reflectance values in heterogeneous and changinglandscapes by using remotely sensed reflectance trends between twodates and spectral unmixing theory (Zhu et al., 2010). We selectedESTARFM as the default model, while STARFM was used when twopairs of input data between two dates were not available for ESTARFM.

1 33 65 97 129 161 193 225 257 289 321 353−3.5

−3

−2.5

−2

−1.5

−1

−0.5

0

0.5

1

1.5

Day of year

NE

E (

g C

m−2

day

−1)

CTLandsatEC observedMODIS

Fig. 9.Averaged seasonal variations of 8-daymeanNEE for thewhole study region in 2005by the three methods (top-down by Carbon Tracker (CT), bottom-up based on Landsatand MODIS imagery) and EC observed measurements.

2.3.4. Accuracy assessmentThe performance of the regression-based diagnosed model for NEE

was evaluated using several statistical parameters, the average error(εa), measuring the quality of the constructed regression tree; the rela-tive error (εr), comparing the quality of several regression trees; the co-efficient of determination (r2), measuring the correlation betweenactual and diagnosed values for the relative variables; and the root-mean-square error (RMSE), measuring the differences between thevalues produced by the rule set approach and the EC-measured NEE.The average error, εα, was calculated as (Xiao et al., 2008; Yang,Huang, et al., 2003; Yang, Xian, et al., 2003):

εa ¼1N

XNt¼1

yt−y′t���

��� ð2Þ

where N is the number of the samples used to construct the diagnosedmodel, yt and yt′ are the tth actual and diagnosed value of the relativevariables respectively. Relative error, εr, is expressed by (Xiao et al.,2008; Yang, Huang, et al., 2003; Yang, Xian, et al., 2003):

εr ¼

1N

XNt¼1

yt−y′t���

���

1N

XNt¼1

yt−yj jð3Þ

where y is the diagnosed mean value. It is used to standardize the aver-age error (εa). The RMSE was calculated as

RMSE ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi1N

XNi¼1

yt−y′t� �2

:

vuut ð4Þ

A two-sided t-test is used to determine the statistical significance ofa1, ⋯, a9 from Eq. (1). A p-value b 0.05 is assumed to be significant.

Fig. 10. Estimated NEE at the regional scale in 2005: (a) EC observed, (b) using Landsat data, and (c) Carbon Tracker (CT).

100 D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

3. Results

3.1. Landscape scale NEE comparisons

The regression-based NEE diagnosed model comprises 19 sub-models (Table 2). The input parameters were weighted by the purefootprint for each EC flux site when using Landsat data. Fig. 3 showsthe land cover maps overlaid by the 8-day cumulative footprint con-tours (50%, 80%, 95% and 99%) for each site. For most sites, the 8-dayfootprint PDFs are distributed asymmetrically around the tower. Thefootprint areas and spatial distribution patterns varied among the fluxsites. The footprint areas for the forest sites are larger than for grassand cropland sites.

For the modeling method comparison, we used the same flux dataand corresponding MODIS data of our method (Landsat-based model)as input parameters for the method of Xiao et al. (2008) (MODIS-based model). The importance of each input parameter, which givesthe percentage of times where each parameters was used in the diag-nosed model, for our method can be quantified as: 33.5% (R1), 42.0%(R2), 38.0% (R3), 33.0% (R4), 45.5% (R5), 49.5% (R7), 79.5% (EVI), 36.0%(NDWI), 51.0% (LST), and 48.5% (land cover) respectively. Compared

with the method using MODIS data, the estimation of NEE using fusedLandsat showed better agreement with the EC flux measurements(Table 3).

As shown in Fig. 4, in terms of land cover, the diagnosedNEE of crop-land has larger scatter, especially when using Landsat data. Both theLandsat- and MODIS-based models can reasonably diagnose landscapeNEE (r2 N 0.59, p b 0.0001); however, the MODIS-based model tendsto underestimate the magnitude of summed NEE, while the Landsat-based model tends to overestimate it, especially for the cropland(Fig. 5). The overestimation from Landsat-based model is likely attri-bute to a large overestimation for croplands due to probable ill-parameterizations for this plant function type as which only based ona limited number of available eddy covariance sites. Model accuracyvaried widely (Table 4). We also compared the observed and estimated8-day NEE in a more detailed manner for each EC flux site (Figs. 6–7,Table 4). For most sites, the performance of the model using Landsatwas better than using MODIS except for the CA-Ojp site (r2 = 0.49 vs.0.57 and RMSE = 0.42 vs. 0.39 g C m−2 day−1 for Landsat vs. MODIS,respectively) and the CA-Ca1 site (r2 = 0.17 vs. 0.24 and RMSE = 1.50vs. 1.43 g C m−2 day−1, Landsat versus MODIS, respectively). For theUS-FPe site, both models did not perform well (r2 = 0.0016 vs. 0.06,

101D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

RMSE = 1.13 vs. 1.16 g C m−2 day−1 for MODIS vs. Landsat, respec-tively). The MODIS-based model of Xiao et al. (2008) also performedpoorly for this site.

A large overestimation in magnitude of NEE by the model was foundat theMead-irrigated continuousmaize site (US-Ne1), theMead-irrigatedmaize-soybean rotation site (US-Ne2), and the Mead-rain-fed maize-soybean rotation site (US-Ne3) using Landsat data, and the US-Ne1 siteusing MODIS data; whereas a large underestimation in magnitude ofNEE occurred at the Audubon Research Ranch (US-Aud), SK-Old Aspen(CA-Oas) using Landsat data, and US-Aud, US-Ne3 and CA-Oas usingMODIS data. For most flux sites, the NEE estimations using Landsat andMODIS both captured most features of observed NEE at seasonal and an-nual scales from 2005 to 2006. Underestimation and overestimation oc-curred for some sites, but there were larger over- or under-estimationsusing MODIS data, such as CA-Oas, US-UMB, US-WCr, US-Bo1, and US-Aud, compared to using Landsat data, such as CA-Ca1, US-Bo1, US-Ne1,US-Ne2, andUS-Ne3 (Fig. 7). Specifically, themodel could not capture ex-ceptionally high or low NEE values (extremes of large carbon release oruptake rates occurred at several sites, such as US-FPe, US-Aud, and CA-Ca1). In general, the model using Landsat data performed better thanwhen using MODIS data. Fig. 8 compares the estimated annual averageNEE in 2005 obtained using the model with Landsat and MODIS datawith EC tower observed values. The annual NEE of cropland was largerthan other land cover types. The estimated annual NEE using Landsatwas larger than MODIS for most EC flux tower sites. Two grassland sites(US-FPe, US-Aud) were carbon sources according to model simulationusing Landsat and MODIS data while EC measurements indicated theywere carbon sinks.

3.2. Regional scale NEE comparisons

The temporal and spatial comparisons of NEE at regional scale wereshown in Figs. 9 and 10. As shown in Fig. 9, compared to the annual var-iation of NEE estimation using a top-down method, such as CarbonTracker (CT) (Peters et al., 2007; Zhang et al., 2013), the differences be-tween CT-derived NEE estimates and values from the two bottom-upupscaling models were large in this study. The time interval of CT wasscaled to eight days from weekly periods using linear interpolation.For the other three methods (EC-Observed, bottom-up based onLandsat and MODIS imagery), the mean NEE was the weighted averageof NEE using the percentage of each land cover. The peak magnitude ofestimated NEE from CT appeared around June 18 (day of year: 169),while for the other two models (Landsat- and MODIS-based models)and measurements it appeared around July 4 (day of year: 185 forLandsat-based model), and August 5 (day of year: 217, for EC-Observed method and for the MODIS-based model).

The regional NEE estimation can be obtained from the analysismethods that upscale data from selected fifteen flux sites with the as-sumption that these sites for different land cover types are representa-tive of the study region (Fig. 10). The regional NEE for each land coverwas estimated using themean values of corresponding site data. The re-gional NEE mapping shows that the strong net carbon uptake in crop-land is distributed in the central to the southeast of the study area.The spatially distributed pattern of NEE produced by CT was similar tothat of the two bottom-up approaches; the largest carbon sink for ex-ample was distributed in cropland land (compare Fig. 10 with Fig. 1).

4. Discussions

There are two advantages to our method. Our model can be appliedto estimating seasonal and annual variations at the landscape scale. Sev-eral studies have used remote-sensing-basedmethods to estimate NEE;however, either the spatial resolution is coarse or the other input datanot derived from remote sensing are not easy to obtain (Maselli et al.,2010; Xiao et al., 2008). For example, using the Carnegie-Stanford-Ames (CASA) biogeochemical model with Landsat imagery, Advanced

Very High Resolution Radiometer (AVHRR) data, forest inventory data,Masek and Collatz (2006) estimated forest carbon fluxes in a disturbedlandscape; however, the Landsat data were not always available. Al-though the revisit cycle of Landsat is 16 days, the available imagerydata frequency was reduced due to cloud contamination. We used animage fusion method to reproduce Landsat-like reflectance data tomake up for the missing periods. Theoretically, it is feasible to producedaily Landsat-like reflectance; but the Landsat-like reflectance oncloudy days cannot be used, and it is very time consuming to producedaily reflectance data.

The PDFs of the integrated flux within EC tower-flux measurementfootprint areas were taken into account in our methods. The size andspatial structure of the footprint obtained using SAFE-f was variable at8-day, monthly, seasonal and annual intervals, and asymmetrically dis-tributed around the tower (Chen, Black, Coops, Hilker, et al., 2009).Consequently, the remotely sensed NEE using Landsat (30-m) orMODIS (1-km) data cannot simply be comparedwith theNEEmeasuredfrom EC tower flux because of their spatial mismatching of their areas.Compared to the simple averaged method using MODIS (Xiao et al.,2008), the estimated NEE using the model with footprint weightedinput parameters was more reasonable.

Compared to the annual variation of NEE estimation using CT meth-od, the peak magnitude of estimated NEE from other three methodwas different (Fig. 9). There exists an apparent seasonal phase shift be-tween the top-down and the bottom-up methods and the main differ-ences in magnitudes between CT and the other three methods(Landsat-, MODIS-basedmodels andmeasurements) were found duringthe peak growing season (July to August). Somemean disagreement be-tween CT and the bottom-up methods can be explained by the differ-ences in small and persistent fluxes in the non-growing seasons (Desai,Helliker, Moorcroft, Andrews, & Berry, 2010). The large overestimationin magnitudes for cropland contributed mainly to the overestimated re-gional carbon sink by the bottom-up upscaling models especially whenusing Landsat data.

The spatially distributed pattern of strong net carbon uptake in crop-land was similar to the pattern found by Hayes et al. (2012). The diag-nosed regional NEE using Landsat had large biases in the summercompared to the EC measurements because of the overestimation offlux from cropland. To sum up, the estimated total annual NEE over thelarge study region (with an area of 1.56 × 109 ha) using the threemodelsthat were being evaluated using the measurements was −3.62 Pg C(Landsat), −2.10 Pg C (MODIS), −0.40 Pg C (CT), −3.56 Pg C (EC-Ob-served), respectively. Themagnitude of the CT estimate was significantlylower than the bottom-upmethods and EC fluxmeasurements. The 20thcentury carbon balance estimation for North America by Hayes et al.(2012) shows the magnitude of CO2 uptake varying from 0.1 to2.0 Pg yr−1 in the 20th century to 1.8 Pg yr−1 in the early 21st century.The reason of the large biases for the two bottom-upmethods (upscalingusing Landsat and MODIS) and EC observed method, perhaps is causedby a too coarse land cover type classification (only 4 groups) and the lim-ited number of flux sites. The dominant land cover type surrounding theEC towers of US-Ne1, US-Ne2 and US-Ne3 is cropland, but the three sitesare closely located so that the EC tower representativeness is poor overthe large cropland area (about 2.6 × 108 ha) in the region.

There are several limitations and constraints while using the NEE di-agnosedmodel in this study. First, the land cover around the flux towerwe used is heterogeneous and our assumption of only one land covertype within the footprint area affects the estimation accuracy. Chenet al. (2012) studied the representativeness of flux tower EC measure-ments for themain sites of Fluxnet-Canada and found that the percent-age of target land cover type area observed by an EC flux tower washigher than 60%. The bias between estimated and observed NEE willbe smaller if the heterogeneous land cover around an EC flux towerwas considered. When the target and other land cover types observedby EC flux tower are taken into account, the performance of NEE diag-nosed model will be better. Second, the phenology may change within

102 D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

the 8-day period invalidating the assumption of representativeness ofLandsat reflectance and LST data for 8 days. For LST data, althoughLandsat samples a local study area around noon time, there can be alarge variation and possible bias (Li et al., 2013). The diagnosed NEE isalso affected when the number of available original Landsat imageswere few. For example, the available original Landsat data for the CA-Ca1 site during the period 2005–2006 was limited with only ten avail-able original data points due to cloud contamination. This would affectthe quality of Landsat-like reflectance. Subsequently, the performanceof diagnosed NEE for this site using Landsat was the worst of all sites(Table 4). For the CA-Ojp site, the original EC observed NEE had manyfluctuations in magnitude in 2005 and the performance for this sitewas also poor. Third, the land-cover type classifications are coarsewith only four major groups. The growing conditions and land-covertypes can vary greatly over large region; few EC flux sites may contrib-ute the bias of NEE diagnosedmodel for the spatial distribution and sea-sonal variations. We selected sixteen EC flux sites, but the performanceof the diagnosedmodel is expected to be better if thereweremore avail-able representativeflux sites to account for land surface heterogeneities.

5. Conclusions

This study introduced a modeling approach for landscape NEE esti-mation at high spatial–temporal resolutions (30-m and 8-day intervals)based on sixteen EC flux-tower measurements and related remotelysensed data. Compared to the similar model of Xiao et al. (2008)which scaled the EC flux tower's NEE at 1-km resolution, we foundthat higher spatial resolution remote sensing data (Landsat vs.MODIS) jointly with flux tower footprint analysis result in bettermodel fit to observed data and thus obtain more robust landscape NEEestimation. The differences in regional NEE estimates between basedon bottom-up up-scaling models and based on top-down inversionmodels (e.g. CT) can be decrease if more flux towers and accurate landcover data are taken into account. This study demonstrated that thedata-driven satellite-based NEE simulation model has the potential toupscale EC flux NEE observations to landscape and regional scaleswith high spatial-temporal resolutions.

Acknowledgments

This research is supported by a research grant (2010CB950704)under the Global Change Program of the Chinese Ministry of Scienceand Technology, a research grant (2012ZD010) of Key Project for theStrategic Science Plan in IGSNRR, CAS, the research grants (41071059& 41271116) funded by the National Science Foundation of China, a Re-search Plan of LREIS (O88RA900KA), CAS, and “One Hundred Talents”program funded by the Chinese Academy of Sciences. Flux data collec-tion in the sites was funded by: US Department of Energy grants: DE-SC0006708; US National Science Foundation grants: DEB-0911461;Canadian Foundation for Climate and Atmospheric Sciences (CFCAS)andNatural Science and Engineering Council of Canada (NSERC) thoughgrants supporting Fluxnet Canada and the Canadian Carbon Program.We thank the USGS EROS data center for providing free Landsat dataand the LP-DAAC and MODIS science team for providing free MODISproducts. Additional contributions from the many researchers involvedin data collection, as in the Fluxnet-Canada and AmeriFlux research net-work, and in-kind support frommany government and private agenciesfor each study site are also gratefully acknowledged.

References

Asrar, G., Fuchs, M., Kanemasu, E. T., & Hatfield, J. L. (1984). Estimating absorbed photo-synthetic radiation and leaf area index from spectral reflectance in wheat1.Agronomy Journal, 76, 300–306.

Barr, A. G., Black, T. A., Hogg, E. H., Griffis, T. J., Morgenstern, K., Kljun, N., et al. (2007). Cli-matic controls on the carbon and water balances of a boreal aspen forest, 1994–2003.Global Change Biology, 13, 561–576.

Barr, A. G., Black, T. A., Hogg, E. H., Kljun, N., Morgenstern, K., & Nesic, Z. (2004).Inter-annual variability in the leaf area index of a boreal aspen–hazelnut forest in re-lation to net ecosystem production. Agricultural and Forest Meteorology, 126, 237–255.

Barr, A. G., Morgenstern, K., Black, T. A., McCaughey, J. H., & Nesic, Z. (2006). Surface en-ergy balance closure by the eddy-covariance method above three boreal forest standsand implications for the measurement of the CO2 flux. Agricultural and ForestMeteorology, 140, 322–337.

Chen, B. Z., Black, T. A., Coops, N. C., Hilker, T., Trofymow, J. A., & Morgenstern, K. (2009).Assessing tower flux footprint climatology and scaling between remotely sensed andeddy covariance measurements. Boundary-Layer Meteorology, 130, 137–167.

Chen, B. Z., Black, T. A., Coops, N. C., Krishnan, P., Jassal, R., Brummer, C., et al. (2009). Sea-sonal controls on interannual variability in carbon dioxide exchange of a near-end-ofrotation Douglas-fir stand in the Pacific Northwest, 1997–2006. Global ChangeBiology, 15, 1962–1981.

Chen, B. Z., Chen, J.M.,Mo, G., Black, A., &Worthy,D. E. J. (2008). Comparison of regional car-bon flux estimates from CO2 concentration measurements and remote sensing basedfootprint integration. Global Biogeochemical Cycles, 22, GB2012. http://dx.doi.org/10.1029/2007GB003024.

Chen, B., Coops, N. C., Fu, D., Margolis, H. A., Amiro, B.D., Black, T. A., et al. (2012). Charac-terizing spatial representativeness of flux tower eddy-covariance measurementsacross the Canadian Carbon Program Network using remote sensing and footprintanalysis. Remote Sensing of Environment, 124, 742–755.

Chen, B., Ge, Q., Fu, D., Yu, G., Sun, X., Wang, S., et al. (2010). A data-model fusion approachfor upscaling gross ecosystem productivity to the landscape scale based on remotesensing and flux footprint modelling. Biogeosciences, 7, 2943–2958.

Choi, M. (2006). A new intensity-hue-saturation fusion approach to image fusion with atradeoff parameter. IEEE Transactions on Geoscience and Remote Sensing, 44, 1672–1682.

Choudhury, B. J. (1987). Relationships between vegetation indices, radiation absorption,and net photosynthesis evaluated by a sensitivity analysis. Remote Sensing ofEnvironment, 22, 209–233.

Churkina, G., Schimel, D., Braswell, B. H., & Xiao, X. (2005). Spatial analysis of growing sea-son length control over net ecosystemexchange.Global Change Biology, 11, 1777–1787.

Ciais, P., Canadell, J. G., Luyssaert, S., Chevallier, F., Shvidenko, A., Poussi, Z., et al. (2010).Can we reconcile atmospheric estimates of the Northern terrestrial carbon sinkwith land-based accounting? Current Opinion in Environmental Sustainability, 2,225–230.

Cook, B.D., Davis, K. J., Wang, W. G., Desai, A., Berger, B. W., Teclaw, R. M., et al. (2004).Carbon exchange and venting anomalies in an upland deciduous forest in northernWisconsin, USA. Agricultural and Forest Meteorology, 126, 271–295.

Deng, F., Chen, J. M., Ishizawa, M., Yuen, C. -W., Mo, G., Higuchi, K. A. Z., et al. (2007). Glob-al monthly CO2 flux inversion with a focus over North America. Tellus SeriesB-Chemical and Physical Meteorology, 59, 179–190.

Desai, A.R., Helliker, B. R., Moorcroft, P. R., Andrews, A. E., & Berry, J. A. (2010). Climaticcontrols of interannual variability in regional carbon fluxes from top-down andbottom-up perspectives. Journal of Geophysical Research – Biogeosciences, 115.

Dragoni, D., Schmid, H. P., Grimmond, C. S. B., & Loescher, H. W. (2007). Uncertaintyof annual net ecosystem productivity estimated using eddy covariance fluxmeasurements. Journal of Geophysical Research-Atmospheres, 112, D17102.http://dx.doi.org/10.1029/2006JD008149.

Dunn, A. L., Barford, C. C., Wofsy, S.C., Goulden, M. L., & Daube, B. C. (2007). A long-termrecord of carbon exchange in a boreal black spruce forest: means, responses to inter-annual variability, and decadal trends. Global Change Biology, 13, 577–590.

Friedl, M.A., McIver, D. K., Hodges, J. C. F., Zhang, X. Y., Muchoney, D., Strahler, A. H., et al.(2002). Global land cover mapping fromMODIS: algorithms and early results. RemoteSensing of Environment, 83, 287–302.

Fry, J. A., Xian, G., Jin, S., Dewitz, J. A., Homer, C. G., Limin, Y., et al. (2011). Completion ofthe 2006 national land cover database for the conterminous United States.Photogrammetric Engineering and Remote Sensing, 77, 858–864.

Gao, B. C. (1996). NDWI — A normalized difference water index for remote sensing ofvegetation liquid water from space. Remote Sensing of Environment, 58, 257–266.

Gao, F., Masek, J., Schwaller, M., & Hall, F. (2006). On the blending of the Landsat andMODIS surface reflectance: Predicting daily Landsat surface reflectance. IEEETransactions on Geoscience and Remote Sensing, 44, 2207–2218.

Gao, F., Masek, J. G., & Wolfe, R. E. (2009). Automated registration and orthorectificationpackage for Landsat and Landsat-like data processing. Journal of Applied RemoteSensing, 3 (033515-033515).

Gourdji, S. M., Mueller, K. L., Schaefer, K., & Michalak, A.M. (2008). Global monthly aver-aged CO2 fluxes recovered using a geostatistical inverse modeling approach: 2. Re-sults including auxiliary environmental data. Journal of Geophysical Research,[Atmospheres], 113, D21115. http://dx.doi.org/10.1029/2007JD009733.

Gu, L., Falge, E. M., Boden, T., Baldocchi, D.D., Black, T. A., Saleska, S. R., et al. (2005). Objec-tive threshold determination for nighttime eddy flux filtering. Agricultural and ForestMeteorology, 128, 179–197.

Gurney, K. R., Baker, D., Rayner, P., & Denning, S. (2008). Interannual variations incontinental-scale net carbon exchange and sensitivity to observing networks esti-mated from atmospheric CO2 inversions for the period 1980 to 2005. GlobalBiogeochemical Cycles, 22, GB3025. http://dx.doi.org/10.1029/2007GB003082.

Gurney, K. R., Law, R. M., Denning, A. S., Rayner, P. J., Baker, D., Bousquet, P., et al. (2003).TransCom 3 CO2 inversion intercomparison: 1. Annual mean control results and sen-sitivity to transport and prior flux information. Tellus Series B-Chemical and PhysicalMeteorology, 55, 555–579.

Gurney, K. R., Law, R. M., Denning, A. S., Rayner, P. J., Baker, D., Bousquet, P., et al. (2002).Towards robust regional estimates of CO2 sources and sinks using atmospheric trans-port models. Nature, 415, 626–630.

Hayes, D. J., Turner, D. P., Stinson, G., McGuire, A.D.,Wei, Y.,West, T. O., et al. (2012). Recon-ciling estimates of the contemporaryNorth American carbon balance among terrestrial

103D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

biospheremodels, atmospheric inversions, and a newapproach for estimating net eco-system exchange from inventory‐based data. Global Change Biology, 18, 1282–1299.

Hollinger, S. E., Bernacchi, C. J., & Meyers, T. P. (2005). Carbon budget of mature no-tillecosystem in North Central Region of the United States. Agricultural and ForestMeteorology, 130, 59–69.

Huete, A., Didan, K., Miura, T., Rodriguez, E. P., Gao, X., & Ferreira, L. G. (2002). Overview ofthe radiometric and biophysical performance of the MODIS vegetation indices.Remote Sensing of Environment, 83, 195–213.

Huete, A.R., Liu, H. Q., Batchily, K., & vanLeeuwen, W. (1997). A comparison of vegetationindices global set of TM images for EOS-MODIS. Remote Sensing of Environment, 59,440–451.

Irish, R. R. (2000). Landsat 7 automatic cloud cover assessment.Algorithms forMultispectral,Hyperspectral, and Ultraspectral Imagery Vi, 4049. (pp. 348–355).

Irish, R. R., Barker, J. L., Goward, S. N., & Arvidson, T. (2006). Characterization of the Landsat-7ETM+ automated cloud-cover assessment (ACCA) algorithm. PhotogrammetricEngineering and Remote Sensing, 72, 1179–1188.

Jackson, T. J., Chen, D., Cosh, M., Li, F., Anderson, M., Walthall, C., et al. (2004). Vegetationwater content mapping using Landsat data derived normalized difference waterindex for corn and soybeans. Remote Sensing of Environment, 92, 475–482.

Keenan, T. F., Baker, I., Barr, A., Ciais, P., Davis, K., Dietze, M., et al. (2012). Terrestrial bio-sphere model performance for inter-annual variability of land-atmosphere CO2 ex-change. Global Change Biology, 18, 1971–1987.

Keenan, T. F., Davidson, E., Moffat, A., Munger, W., & Richardson, A.D. (2012). Usingmodel-data fusion to interpret past trends, and quantify uncertainties in future projec-tions, of terrestrial ecosystem carbon cycling. Global Change Biology, 18, 2555–2569.

LeMone, M.A., Grossman, R. L., McMillen, R. T., Liou, K. N., Ou, S.C., Mckeen, S., et al. (2002).Cases-97: Late-morning warming and moistening of the convective boundary layerover the Walnut River Watershed. Boundary-Layer Meteorology, 104, 1–52.

Li, H., Manjunath, B.S., & Mitra, S. K. (1995). Multisensor image fusion using the wavelettransform. Graphical Models and Image Processing MyNotes, 57, 235–245.

Li, Z. -L., Tang, B. -H., Wu, H., Ren, H., Yan, G., Wan, Z., et al. (2013). Satellite-derived landsurface temperature: Current status and perspectives. Remote Sensing of Environment,131, 14–37.

Mahadevan, P., Wofsy, S.C., Matross, D.M., Xiao, X. M., Dunn, A. L., Lin, J. C., et al. (2008). Asatellite-based biosphere parameterization for net ecosystem CO2 exchange: Vegeta-tion Photosynthesis and Respiration Model (VPRM). Global Biogeochemical Cycles, 22,GB2005. http://dx.doi.org/10.1029/2006GB002735.

Masek, J. G., & Collatz, G. J. (2006). Estimating forest carbon fluxes in a disturbedsoutheastern landscape: Integration of remote sensing, forest inventory, andbiogeochemical modeling. Journal of Geophysical Research – Biogeosciences, 111,G01006. http://dx.doi.org/10.1029/2005JG000062.

Masek, J. G., Vermote, E. F., Saleous, N. E., Wolfe, R., Hall, F. G., Huemmrich, K. F., et al.(2006). A Landsat surface reflectance dataset for North America, 1990–2000. IEEEGeoscience and Remote Sensing Letters, 3, 68–72.

Maselli, F., Chiesi, M., Fibbi, L., & Moriondo, M. (2008). Integration of remote sensing andecosystem modelling techniques to estimate forest net carbon uptake. InternationalJournal of Remote Sensing, 29, 2437–2443.

Maselli, F., Gioli, B., Chiesi, M., Vaccari, F., Zaldei, A., Fibbi, L., et al. (2010). Validating anintegrated strategy to model net land carbon exchange against aircraft flux measure-ments. Remote Sensing of Environment, 114, 1108–1116.

Metwalli,M. R., Nasr, A.H., Allah, O. S. F., El-Rabaie, S., & Abd El-Samie, F. E. (2010). Satelliteimage fusion based on principal component analysis and high-pass filtering. Journal ofthe Optical Society of America A-Optics Image Science and Vision, 27, 1385–1394.

Michalak, A.M., Bruhwiler, L., & Tans, P. P. (2004). A geostatistical approach to surface fluxestimation of atmospheric trace gases. Journal of Geophysical Research: Atmospheres,109, D14109. http://dx.doi.org/10.1029/2003JD004422.

Moffat, A.M., Papale, D., Reichstein, M., Hollinger, D. Y., Richardson, A.D., Barr, A. G., et al.(2007). Comprehensive comparison of gap-filling techniques for eddy covariance netcarbon fluxes. Agricultural and Forest Meteorology, 147, 209–232.

Mueller, K. L., Gourdji, S. M., & Michalak, A.M. (2008). Global monthly averaged CO2fluxes recovered using a geostatistical inverse modeling approach: 1. Results usingatmospheric measurements. Journal of Geophysical Research: Atmospheres, 113,D21114. http://dx.doi.org/10.1029/2007JD009734.

Myneni, R. B., Hall, F. G., Sellers, P. J., & Marshak, A. L. (1995). The interpretation of spectralvegetation indexes. IEEE Transactions on Geoscience and Remote Sensing, 33, 481–486.

Naidu, V. P.S., & Raol, J. R. (2008). Pixel-level image fusion using wavelets and principalcomponent analysis. Defence Science Journal, 58, 338–352.

Nave, L. E., Gough, C. M., Maurer, K. D., Bohrer, G., Hardiman, B.S., Le Moine, J., et al.(2011). Disturbance and the resilience of coupled carbon and nitrogen cycling in anorth temperate forest. Journal of Geophysical Research – Biogeosciences, 116,G04016. http://dx.doi.org/10.1029/2011JG001758.

Nunez, J., Otazu, X., Fors, O., Prades, A., Pala, V., & Arbiol, R. (1999). Multiresolution-basedimage fusion with additive wavelet decomposition. IEEE Transactions on Geoscienceand Remote Sensing, 37, 1204–1211.

Papale, D., & Valentini, A. (2003). A new assessment of European forests carbon ex-changes by eddy fluxes and artificial neural network spatialization. Global ChangeBiology, 9, 525–535.

Penuelas, J., Gamon, J. A., Griffin, K. L., & Field, C. B. (1993). Assessing community type,plant biomass, pigment composition, and photosynthetic efficiency of aquatic vegeta-tion from spectral reflectance. Remote Sensing of Environment, 46, 110–118.

Penuelas, J., Garbulsky, M. F., & Filella, I. (2011). Photochemical reflectance index (PRI)and remote sensing of plant CO2 uptake. New Phytologist, 191, 596–599.

Penuelas, J., Garbulsky, M. F., Gamon, J., Inoue, Y., & Filella, I. (2011). The photochemicalreflectance index (PRI) and the remote sensing of leaf, canopy and ecosystem radia-tion use efficiencies: A review andmeta-analysis. Remote Sensing of Environment, 115,281–297.

Peters, W., Jacobson, A.R., Sweeney, C., Andrews, A. E., Conway, T. J., Masarie, K., et al.(2007). An atmospheric perspective on North American carbon dioxide ex-change: CarbonTracker. Proceedings of the National Academy of Sciences, 104,18925–18930.

Peters, W., Krol, M. C., van der Werf, G. R., Houweling, S., Jones, C. D., Hughes, J., et al.(2010). Seven years of recent European net terrestrial carbon dioxide exchangeconstrained by atmospheric observations. Global Change Biology, 16, 1317–1337.

Peylin, P., Bousquet, P., Le Quere, C., Sitch, S., Friedlingstein, P., McKinley, G., et al. (2005).Multiple constraints on regional CO2 flux variations over land and oceans. GlobalBiogeochemical Cycles, 19, GB1011. http://dx.doi.org/10.1029/2003GB002214.

Piao, S. L., Fang, J. Y., Ciais, P., Peylin, P., Huang, Y., Sitch, S., et al. (2009). The carbon bal-ance of terrestrial ecosystems in China. Nature, 458, 1009–1013.

Quinlan, J. R. (1992). Learning with continuous classes. Proceedings of the 5th AustralianJoint Conference On Articial Intelligence (pp. 343–348).

Rahman, A., Sims, D., Cordova, V., & El-Masri, B. (2005). Potential of MODIS EVI and sur-face temperature for directly estimating per-pixel ecosystem C fluxes. GeophysicalResearch Letters, 32, L19404.

Ranson, K. J., Daughtry, C. S. T., Biehl, L. L., & Bauer, M. E. (1985). Sun-view angle effects onreflectance factors of corn canopies. Remote Sensing of Environment, 18, 147–161.

Raupach, M. R. (1994). Simplified expressions for vegetation roughness length andzero-plane displacement as functions of canopy height and area index. Boundary-LayerMeteorology, 71, 211–216.

Reichstein, M., Falge, E., Baldocchi, D., Papale, D., Aubinet, M., Berbigier, P., et al. (2005).On the separation of net ecosystem exchange into assimilation and ecosystem respi-ration: review and improved algorithm. Global Change Biology, 11, 1424–1439.

Roujean, J. L., Leroy, M., & Deschamps, P. Y. (1992). A bidirectional reflectance model ofthe Earth's surface for the correction of remote-sensing data. Journal of GeophysicalResearch-Atmospheres, 97, 20455–20468.

RuleQuest (2012). http://www.rulequest.com/ (last visited on Nov. 15, 2012)Ryu, Y., Baldocchi, D.D., Ma, S., & Hehn, T. (2008). Interannual variability of evapotranspira-

tion and energy exchange over an annual grassland in California. Journal of GeophysicalResearch-Atmospheres, 113, D09104. http://dx.doi.org/10.1029/2007JD009263.

Sellers, P. J., Berry, J. A., Collatz, G. J., Field, C. B., & Hall, F. G. (1992). Canopy reflectance,photosynthesis, and transpiration. III. A reanalysis using improved leaf models anda new canopy integration scheme. Remote Sensing of Environment, 42, 187–216.

Sobrino, J. A., Jiménez-Muñoz, J. C., & Paolini, L. (2004). Land surface temperature retrievalfrom LANDSAT TM 5. Remote Sensing of Environment, 90, 434–440.

Stinson, G., Kurz, W. A., Smyth, C. E., Neilson, E. T., Dymond, C. C., Metsaranta, J. M., et al.(2011). An inventory-based analysis of Canada's managed forest carbon dynamics,1990 to 2008. Global Change Biology, 17, 2227–2244.

Suyker, A. E., & Verma, S. B. (2008). Interannual water vapor and energy exchange in an ir-rigatedmaize-based agroecosystem.Agricultural and ForestMeteorology, 148, 417–427.

Tang, X., Wang, Z., Liu, D., Song, K., Jia, M., Dong, Z., et al. (2012). Estimating the net eco-system exchange for the major forests in the northern United States by integratingMODIS and AmeriFlux data. Agricultural and Forest Meteorology, 156, 75–84.

Tu, T. M., Su, S.C., Shyu, H. C., & Huang, P.S. (2001). Efficient intensity-hue-saturation-basedimage fusion with saturation compensation. Optical Engineering, 40, 720–728.

Turner, D. P., Urbanski, S., Bremer, D., Wofsy, S.C., Meyers, T., Gower, S. T., et al. (2003). Across-biome comparison of daily light use efficiency for gross primary production.Global Change Biology, 9, 383–395.

Veroustraete, F., Patyn, J., & Myneni, R. (1996). Estimating net ecosystem exchange of car-bon using the normalized difference vegetation index and an ecosystem model.Remote Sensing of Environment, 58, 115–130.

Wickham, J.D., Stehman, S. V., Gass, L., Dewitz, J., Fry, J. A., & Wade, T. G. (2013). Accuracyassessment of NLCD 2006 land cover and impervious surface. Remote Sensing ofEnvironment, 130, 294–304.

Wilson, T., & Meyers, T. (2007). Determining vegetation indices from solar and pho-tosynthetically active radiation fluxes. Agricultural and Forest Meteorology, 144,160–179.

Wulder, M.A., Dechka, J. A., Gillis, M.A., Luther, J. E., Hall, R. J., & Beaudoin, A. (2003). Op-erational mapping of the land cover of the forested area of Canada with Landsat data:EOSD land cover program. Forestry Chronicle, 79, 1075–1083.

Wulder, M.A., White, J. C., Goward, S. N., Masek, J. G., Irons, J. R., Herold, M., et al. (2008).Landsat continuity: Issues and opportunities for land cover monitoring. RemoteSensing of Environment, 112, 955–969.

Xian, G., Homer, C., & Fry, J. (2009). Updating the 2001 National Land Cover Database landcover classification to 2006 by using Landsat imagery change detection methods.Remote Sensing of Environment, 113, 1133–1147.

Xiao, J. F., Chen, J. Q., Davis, K. J., & Reichstein, M. (2012). Advances in upscaling of eddycovariance measurements of carbon and water fluxes. Journal of GeophysicalResearch – Biogeosciences, 117, G00J01. http://dx.doi.org/10.1029/2011JG001889.

Xiao, J. F., Davis, K. J., Urban, N. M., Keller, K., & Saliendra, N. Z. (2011). Upscaling carbonfluxes from towers to the regional scale: Influence of parameter variability and landcover representation on regional flux estimates. Journal of Geophysical Research –Biogeosciences, 116, G00J06. http://dx.doi.org/10.1029/2010JG001568.

Xiao, X. M., Hollinger, D., Aber, J., Goltz, M., Davidson, E. A., Zhang, Q. Y., et al. (2004).Satellite-based modeling of gross primary production in an evergreen needleleaf for-est. Remote Sensing of Environment, 89, 519–534.

Xiao, J. F., Zhuang, Q. L., Baldocchi, D.D., Law, B. E., Richardson, A.D., Chen, J. Q., et al. (2008).Estimation of net ecosystem carbon exchange for the conterminous United States bycombining MODIS and AmeriFlux data. Agricultural and Forest Meteorology, 148,1827–1847.

Xiao, J., Zhuang, Q., Law, B. E., Baldocchi, D.D., Chen, J., Richardson, A.D., et al. (2011).Assessing net ecosystem carbon exchange of U.S. terrestrial ecosystems by integrat-ing eddy covariance flux measurements and satellite observations. Agricultural andForest Meteorology, 151, 60–69.

104 D. Fu et al. / Remote Sensing of Environment 141 (2014) 90–104

Yang, L., Huang, C., Homer, C. G., Wylie, B. K., & Coan, M. J. (2003). An approach formapping large-area impervious surfaces: synergistic use of Landsat-7 ETM+and high spatial resolution imagery. Canadian Journal of Remote Sensing, 29,230–240.

Yang, L., Xian, G., Klaver, J. M., & Deal, B. (2003). Urban land-cover change detection throughsub-pixel imperviousness mapping using remotely sensed data. PhotogrammetricEngineering and Remote Sensing, 69, 1003–1010.

Yuan, W., Luo, Y., Li, X., Liu, S., Yu, G., Zhou, T., et al. (2011). Redefinition and global esti-mation of basal ecosystem respiration rate. Global Biogeochemical Cycles, 25, GB4002.http://dx.doi.org/10.1029/2011GB004150.

Zhang, Y. (2004). Understanding image fusion. Photogrammetric Engineering and RemoteSensing, 70, 657–661.

Zhang, H., Chen, B., & Peters, W. (2013). Inversing grided land surface carbon fluxes focus-ing on Asia region. EGU General Assembly Conference Abstracts (pp. 6895).

Zhang, L., Furumi, S.,Muramatsu, K., Fujiwara, N., Daigo,M., & Zhang, L. (2007). A new veg-etation index based on the universal pattern decomposition method. InternationalJournal of Remote Sensing, 28, 107–124.

Zhu, X. L., Chen, J., Gao, F., Chen, X. H., & Masek, J. G. (2010). An enhanced spatial and tem-poral adaptive reflectance fusion model for complex heterogeneous regions. RemoteSensing of Environment, 114, 2610–2623.