BOSTON UNIVERSITY

GRADUATE SCHOOL OF ARTS AND SCIENCES

Thesis

ANALYSIS OF YEAR 2000 MODIS LAI / FPAR PRODUCT

by

SETH W. HOFFMAN

B.A., Boston University, 1996

Submitted in partial fulfillment of the

requirements for the degree of

Master of Arts

2001

Approved by

First Reader ___________________________________________________________

Ranga Myneni, Ph.D. Associate Professor of Geography

Second Reader __________________________________________________________

Yuri Knyazikhin, Ph.D. Research Associate Professor of Geography

Third Reader ___________________________________________________________

Crystal Schaaf, Ph.D. Research Associate Professor of Geography

iii

Acknowledgements

I would like to thank the members of the Department of Geography at Boston University

who have shared their knowledge and views to help expand mine. Particular thanks goes

to Ranga Myneni who has guided my time here, as he does all his students, with great

care and interest. He saw my potential and provided an opportunity to continue an

academic life with both financial and moral support. Without his support I would be

teaching English to city-folk in Japan instead of to our bevy of Chinese students on my

native soil.

Cheers also go out to the other members of the Boston University Climate and

Vegetation Research Group, especially Juri Knyazikhin, whose stream of work carried

my efforts along the river of research. Graham Walker provided excellent counterpoint to

some of the more stress inducing students during my stint as a teaching fellow for

Introduction to Climatology. Tony Soares provided top-notch computer system

administration, and his hard fought campaign to make OpenGL available on the crs

cluster was greatly appreciated. I would like to extend my gratitude to Raymond Dezzani

for his patience and infinitely deep well of reading material. Finally, I would like to thank

Crystal Schaaf for taking her valuable time to be a reader for this thesis.

iv

ANALYSIS OF YEAR 2000 MODIS LAI / FPAR PRODUCT

SETH W. HOFFMAN

ABSTRACT

An algorithm based on the physics of radiative transfer in vegetation canopies for the

retrieval of vegetation green leaf area index (LAI) and fraction of absorbed

photosynthetically active radiation (FPAR) from surface reflectances was developed and

implemented for operational processing of data from the MODerate resolution Imaging

Spectroradiometer (MODIS). Extensive pre-launch validation work was performed to

assure the functionality of the physics built into the algorithm. With the launch of the

Terra platform in December of 1999, post-launch validation work has focused on

ensuring that the product corresponds to physical reality. This work has proceeded with

rigorous comparisons to acquired ground truth data from field campaigns, and through

statistical analysis of the LAI, FPAR and QA data fields. Potential users of the various

levels of MODIS LAI/FPAR product (MOD15) temporal composites should be aware

that even with the positive initial validation results presented here, changes due to

calibration, geolocation, cloud screening and atmospheric correction will be on-going.

v

Contents

1 Introduction 1

2 The MODIS Instrument 3

3 The MODIS LAI / FPAR Algorithm 7

3.1 The Global Biome Map 7

3.2 Input Uncertainties and Solution Distributions 10

3.3 Energy Conservation as a Constraint 12

3.4 Spectral Invariance 13

4 Performance of the MODIS LAI / FPAR Algorithm 14

4.1 Input and Model Uncertainties 14

4.2 NDVI and LAI/FPAR Relations 16

4.3 Impact of Biome Misclassification 18

4.4 Scale Dependence of the Algorithm 19

4.5 Reflectance Saturation 21

5 The MODIS LAI / FPAR Product 24

6 Validation 40

6.1 Bigfoot 42

6.2 SAFARI 2000 Wet Season Campaign 43

6.3 Summer 2000 Field season 46

6.3.1 Instruments 47

6.3.1.1 LAI-2000 47

vi

6.3.1.2 AccuPAR Ceptometer 48

6.3.1.3 ASD Spectroradiometer 49

6.3.1.4 LI-1800 49

6.3.2 Field Campaigns 49

6.3.2.1 Harvard Forest 49

6.3.2.1.1 Grid Experiment 50

6.3.2.1.2 Road Experiment 50

6.3.2.1.3 Short Transect Experiment 51

6.3.2.2 Kejimkujic National Park 51

6.3.2.3 Ruokolahti 52

7 Conclusions 54

8 References 56

vii

List of Tables

1 Description of MODIS spectral bands 4

2 Description of MODIS product processing levels for data from

instruments aboard the Terra satellite 4

3 Translation of U. M.D. land cover types into B.U. biome types 8

4 Effect of biome misclassification on LAI retrievals 19

5 Retrieval indices and mean LAI from algorithm runs using Landsat

and LASUR data and LUTs 21

6 Retrieval Index and frequency of LAI under saturation for POLDER

data 23

7 Summary of the MODIS LAI/FPAR Scientific Data Sets (SDS) 25

8 The FPAR / LAI quality control variable summary chart 30

9 Correlation between average (main – backup) differences and

gradated values of LAI and FPAR produced by the main algorithm

for each difference 33

10 Regression results between main and back-up values for

corresponding pixels from consecutively dated A1 data 33

11 LAI-net sites for validation of LAI/FPAR products from MODIS

and other moderate resolution sensors 41

12 High-resolution IKONOS-based LAI retrievals compared to transect

and grid field measurements at Maun, Botswana 45

viii

List of Figures

1 MODIS 1km land cover map, composited into B.U. biome types 9

2 Illustration of factors used to generate MODIS LAI 10

3 Dependence of the Retrieval Index on uncertainties in

measurements and model simulations 16

4 Red – NIR relations and NDVI / LAI / FPAR relations 17

5 Illustration of the convergence property of the LAI model from

different base soil conditions. 22

6 Global tile coverage of MOD15A2 product for year 2000 26

7 Tiles produced during year 2000 27

8 Color-coded maps of MODIS LAI and FPAR fields for the boreal

summer of year 2000 28

9 Comparison of pre- and post- spatially composited FPAR images of

Australia 29

10 MODIS LAI product for a region in the Amazonia for two periods

in July 2000, along with corresponding QA fields 32

11 Differences between values produced by the main and by the

backup algorithm as a function of values from the main algorithm 35

12 Average QA values as a function of latitude 36

13 Distribution of LAI and FPAR values from the global coverage

8-day period 06/25 to 07/02 37

ix

14 Cumulative distributions of differences between main and backup

values of pixels over consecutive days 38

15 Comparison of magnitudes of LAI and FPAR differences between

main and backup values 39

16 Comparison of MODIS LAI product with SAFARI 2000 wet

season field campaign measurements 44

17 LAI maps of a 5 x 5km region at Maun in Botswana retrieved by

the LAI/FPAR algorithm with MODIS, LANDSAT ETM+ and

IKONOS data 46

18 Distributions of LAI retrievals shown in Fig. 16 together with field

measurements at Maun, Botswana 47

x

List of Acronyms

a(λ) Absorptance at λ

AVHRR Advanced Very High Resolution Radiometer

BRF Bi-directional Reflectance Factor

DAAC Distributed Active Archive Center

ε Uncertainty in spectral reflectance measurement

EOS Earth Observing System

EROS Earth Resources Observation System

ETM+ Enhanced Thematic Mapper +

FPAR Fraction of absorbed Photosynthetically Active Radiation

GLCC Global Land Cover Characterization database

HDF Hierarchical Data Format

IKONOS Commercial 1m resolution satellite

LAI Leaf Area Index

Landsat Land remote sensing Satellite

LASUR Land Surface Reflectance

LTER Long Term Ecological Research network

LUT Look-Up Table

MAUP Modifiable Areal Unit Problem

MISR Multi-angle Imaging Spectrometer

MOD15A1 Daily LAI/FPAR product from MODIS

MOD15A2 8-day composite LAI/FPAR product from MODIS

MOD15A3 Monthly composite LAI/FPAR product from MODIS

MOD12 96-day land cover product from MODIS

MOD35 Global cloud mask product from MODIS

MODAPS MODIS Adaptive Processing System

MODIS Moderate Resolution Imaging Spectroradiometer

MODLAND MODIS LAND science team

NIR Near Infra-Red

xi

NDVI Normalized Differential Vegetation Index

POLDER Polarization and Directionality of Earth’s Reflectances

QA Quality Assessment

RI Retrieval Index

SeaWiFS Sea-viewing Wide Field-of-view Sensor

SLCR Seasonal Land Cover Regions

t(λ) Transmittance at λ

Terra EOS-AM spacecraft (Launched 1999)

1

1 INTRODUCTION

Leaf area index (LAI) and fraction of photosynthetically active radiation (0.4 - 0.7 µm)

absorbed by vegetation (FPAR) characterize vegetation canopy functioning and energy

absorption capacity. LAI is defined as one sided green leaf area per unit ground area in

broadleaf canopies and as the projected needle leaf area in coniferous canopies. LAI and

FPAR are key parameters in most ecosystem productivity models and global models of

climate, hydrology, biogeochemistry and ecology (Sellers et al., 1997). For effective use

in large-scale models, these variables must be collected over a long period of time and

should represent every region of the terrestrial surface. Satellite remote sensing is the

most effective means of collecting such global fields on a regular basis. Advances in

remote sensing technology (Deschamps et al., 1994; Justice et al., 1998; Diner et al.,

1999) and radiative transfer modeling (Myneni and Ross, 1991; Ross et al., 1992) have

improved the possibility of accurate estimation of these parameters from spectral and

angular dimensions of remotely sensed data. The launch of Terra with MODIS

(MODerate resolution Imaging Spectroradiometer) and MISR (Multiangle Imaging

SpectroRadiometer) instruments onboard began a new era in remote sensing of the Earth

system. In contrast to previous instruments with limited angular and spectral capabilities,

MODIS and MISR together allow for rich spectral and angular sampling of the reflected

radiation field. This sets new demands on retrieval techniques for geophysical parameters

in order to take full advantage of these instruments. In this context, a synergistic

algorithm has been developed for the extraction of LAI and FPAR from MODIS- and

2

MISR-measured canopy reflectance data, with the flexibility of using the same algorithm

in MODIS-only and MISR-only modes as well.

In this thesis, an overview of the MODIS LAI/FPAR research is provided. This

includes a description of the salient features of the algorithm, evaluation of its

performance, details of the product from year one of MODIS data and initial results of

validation with field data. The MODIS LAI/FPAR product is currently being produced

operationally and is available to the user community free of charge.

3

2 THE MODIS INSTRUMENT

The Terra (EOS AM-1) satellite, which carries the MODerate Resolution Imaging

Spectroradiometer and four other instruments, was launched on December 18, 1999.

Scientific data collection began on February 24, 2000. The satellite orbits the Earth at an

altitude of 705 km in a sun synchronous polar orbit. Descending southward, Terra crosses

the equator at the same local time every day. A sister satellite, Aqua (EOS PM-1),

carrying a second MODIS instrument will be launched no earlier than December, 2001.

With a viewing swath of 2,330 km, MODIS produces global coverages every one to

two days. A continuously rotating double-sided mirror scans across-track through ± 55º

to direct light into the optical assemblies. The detectors are sensitive to 36 spectral bands

at wave

and are produced, depending on the band, at three different spatial resolutions (Table 1).

Data from the satellite goes through up to 5 different levels of processing (Table 2).

After initial Level 0 processing at the EOS Data and Operations System, the Goddard

Space Flight Center Earth Sciences (GES) DAAC produces Level 1A, 1B, geolocation

and cloud mask products. Higher-level products are produced by the MODIS Adaptive

Processing System (MODAPS).

Building upon and extending the heritage of Landsat and AVHRR, MODIS is ideal for

monitoring large-scale changes in the biosphere. The series of MODIS instruments are

expected to help develop a 15-year data set for comprehensive global change research.

Beyond constructing global dynamical models of Earth’s systems, the set of 44 data

4

products will also be used to help differentiate natural changes from those caused by

human activities. The nature of MODIS observations allow for a multi-disciplinary study

of atmospheric, land and ocean processes and their interactions. In addition to the above

categories a fourth science effort, led by the calibration discipline group, will characterize

and monitor the performance of the instrument.

Bands Resolution (m) Primary Use 1-2 250 Land / Cloud / Aerosol Boundaries 3-7 500 Land / Cloud / Aerosol properties 8-16 1000 Ocean color / Phytoplankton / Biogeochemistry 17-19 1000 Atmospheric water vapor 20-23 1000 Surface/Cloud temperature 24-25 1000 Atmospheric temperature 26-28 1000 Cirrus clouds water vapor 29 1000 Cloud properties 30 1000 Ozone 31-32 1000 Surface/Cloud temperature 33-36 1000 Cloud top altitude

Table 1. Description of MODIS spectral bands. For more details see http://modis.gsfc.nasa.gov/

Level Description 0 Reconstructed from satellite with communication artifacts removed 1 Time referenced and annotated Level 0 data (calibrated) 2 Derived geophysical variables at original resolution and location 3 Variables are mapped onto uniform space-time grids 4 Model output or analysis results from lower level data.

Table 2. Description of the meaning of the different processing levels for data from instruments aboard the Terra satellite. For more details, see http://terra.nasa.gov/Brochure/Sect_5-1.html

The MODIS Atmosphere Discipline Group is charged with developing algorithms to

determine cloud and aerosol properties. These properties play a major role in the global

5

climate and are required for predictive global climate models. Six products are under the

purview of this group, including the daily cloud mask (MOD35) product.

The MODIS Land (MODLAND) team is responsible for the development and

validation of algorithms used to create land cover and land cover change data sets.

Centralized creation of these products removes the burden of data processing from the

user community and ensures a single common data set. The set of 16 MODLAND

products includes surface reflectance, albedo, land cover, vegetation indices, LAI, FPAR,

snow and ice cover, surface temperatures and anomalies, and a number of other

biophysical parameters.

Due to their immense capacity to store heat and carbon, the oceans play a major role in

moderating climate. To what degree oceans influence climate, and do climate changes

ultimately change the oceans are some of the questions to be answered by data from the

MODIS Ocean group. Three categories of products, ocean color, sea surface temperature

and ocean primary production, will be produced to help answer these and other questions.

The harsh environment of outer space, from radiation bombardment to micrometeorite

collisions to extreme temperature shifts, damages all material over time. To ensure the

continuing quality of the MODIS data sets over its mission life, four separate systems are

included on-board the instrument. Each of these systems will provide stable data points

for the calibration group so that they can execute and update calibrations to the incoming

raw data. The Solar Diffuser and Solar Diffuser Stability monitor use the sun’s constant

output to check for changes in system response to incoming light. An internal light source

called the Spectroradiometric Calibration Assembly will provide information on

6

radiometric, spectral and geometric characteristics of MODIS systems. The Blackbody

Assembly will be used to test the performance of infrared detectors. Changing at a rate

slower than the lifetime of the EOS mission, the moon’s surface is another view that

works similar to the Solar Diffuser. Stars are too dim to be registered by the detectors, so,

a zero-point for calibration reference is made by viewing deep space.

7

3 THE MODIS LAI/FPAR ALGORITHM

Inputs to the operational MODIS LAI/FPAR algorithm include a global biome map and

up to seven atmosphere-corrected surface spectral (648, 858, 470, 555, 1240 and 2130

nm) bi-directional reflectance factors (BRFs) and their uncertainties (MOD09 product).

At present, only the first two of these bands (red and NIR) are being used. The outputs

are the most probable values of pixel LAI, FPAR and their respective dispersions. The

theoretical basis of the algorithm is given in Knyazikhin et al. (1998a) and the

implementation aspects are discussed in Knyazikhin et al. (1999). A look-up-table (LUT)

method is used to achieve inversion of the three-dimensional radiative transfer problem.

When this method fails to localize a solution, a back-up method based on relations

between the normalized difference vegetation index (NDVI) and LAI/FPAR is utilized,

together with a biome classification map, to perform the retrievals (Myneni et al., 1997).

The principal features of the LUT-based method are summarized below.

3.1 Global Biome Map

Global vegetation is stratified from the 13 U.M.D (Hansen et al., 2000) vegetated classes

into six canopy architectural types, or biomes, in our LAI/FPAR retrieval method (Table

3). The six biomes are grasses and cereal crops, shrubs, broadleaf crops, savannas,

broadleaf forests and needle leaf forests. These biomes span structural variations along

the horizontal (homogeneous vs. heterogeneous) and vertical (single- vs. multi-story)

8

dimensions, canopy height, leaf type, soil brightness and climate (precipitation and

temperature) space of herbaceous and woody vegetation globally.

B.U. Classes U. M.D Classes Grasses & Cereal Crops Grass Shrubs Open Shrub-land

Broadleaf Crops Cropland Crop Mosaic

Savannas Woody Savanna Savanna

Broadleaf Forests

Evergreen Broadleaf Forest Deciduous Broadleaf Forest Mixed Forest Closed Shrub-land Wetland

Needle Forests Evergreen Needle Forest Deciduous Needle Forest

Table 3. Translation of U.M.D land cover types into B.U biome types. Used for the MOD15 product. From Lotsch (1999).

The biome map reduces the number of unknowns in the inverse problem through the

use of simplifying assumptions (e.g., model leaf normal orientation distributions) and

standard constants (e.g., leaf, wood, litter and soil optical properties) that are assumed to

vary with biome and soil types only. This approach is similar to that adopted in many

global models which assume certain key parameters to vary only by vegetation type and

utilize a land cover classification to achieve spatialization.

Translation of the U.M.D land cover classes into the six biomes is ambiguous with

respect to several classes. The only global map source with the requisite details is the

Global Land Cover Characterization Data Base (GLCC) available from EROS (Earth

Resources Observation System) Data Center (Loveland et al., 2000). Seasonal land cover

9

regions (SLCR) provide a framework for aggregation into broader classes, e.g., the six

biomes. To generate a six-biome map, the Hansen et al. (2000) U.M.D classification map

is used together with the SLCR map (Lotsch et al., 2001). A direct assignment is

performed for classes in the U.M.D scheme that can be directly translated into one of the

six biomes. For classes that do not allow a direct translation, the respective SLCR label is

retrieved, and a biome is assigned using an SLCR-biome look-up table. This is performed

on a per-pixel basis. The biome map thus obtained is shown in Fig. 1 and is used by the

at-launch algorithm to produce the data used in the scientific quality analysis discussed

later. The MODIS derived Land Cover Product (MOD12Q1), which is updated every 96

days, will provide a variety of land cover classification schemes, including that used for

the LAI / FPAR product. This will end the need for the translation and aggregation of

other global coverages.

Figure 1. Global 1km six-biome map used by the MODIS LAI/FPAR algorithm. From Myneni et al. (2001).

10

3.2 Input Uncertainties and Solution Distributions

The LAI/FPAR algorithm compares MODIS directional spectral reflectances to

comparable values evaluated from model-based entries stored in a look-up-table (LUT)

and derives the distribution of all possible solutions, i.e., LAI and FPAR distribution

functions. Mean and variance of the distribution functions are archived. This is illustrated

in Fig. 2, where the algorithm inputs are shown in the red and near-infrared space (panel

a) and the outputs as solution distributions (panel b). A one-to-one match between the

inputs and outputs is realized only in the case of error-free inputs, a perfect algorithm and

sufficient information required to uniquely localize a solution (Zhang et al., 2000). For

the operational algorithm this is rarely, if ever, the case.

Figure 2. Illustration of factors used to generate MODIS LAI. (a) Distribution of vegetated pixels with respect to their reflectances at red and near infrared spectral bands derived from SeaWiFS data (September 22, 1997). A point on the red-NIR (near infrared) plane and an area about it (white ellipses defined by a chi-squared distribution) can be treated as the measured BRF and its uncertainty. Each canopy/soil pattern for which modeled reflectances belong to the ellipse shown is an acceptable solution. For each set of observed spectral reflectances and their uncertainties, one defines a cumulative solution distribution function F(l) as the portion of LAI values which are less than l. (b) The solution density distribution function dF(l)/d(l) is shown for five different pixels. The mean LAI over this distribution and its dispersion are taken as the LAI retrieval and its uncertainty, respectively. The labels: 0.1, 1, 2, 3 and 5 indicate LAI values of the selected pixels. The algorithm will report the LAI value of the highest peak. If the distribution is flat across a range of LAI values, the mean value will be reported along with the saturation condition QA flag. From Knyazikhin et al. (1998a).

11

First, two different locations in the input space can correspond to the same value of the

output, but for different biome types. The algorithm should account for differences in the

structure and optics of these biomes in a way that the same value of LAI is retrieved in

both cases. Second, a point in the input space may correspond to multiple LAI values,

because, for example, different combinations of LAI and soil types can result in the same

value of canopy spectral reflectances. This means that the inputs do not contain sufficient

information to localize a unique solution. Third, in the case of a dense canopy, its

reflectance in one or more directions can be insensitive to various parameter values (e.g.,

LAI) characterizing the canopy because, the reflectance of solar radiation from the

underlying soil surface or lower leaf-stories is completely obscured by the upper leaves

(Price, 1993; Liu and Huete, 1995; Jasinski, 1996; Carlson and Ripley, 1997). When this

happens, the canopy reflectance is said to belong to the saturation domain (Gobron et. al.,

1997, Knyazikhin et al., 1998b). The distribution of LAI values will appear flat over this

range of LAI, illustrating that the solutions all have equal probability of occurrence (case

5 Fig. 2b). Therefore, all LAI values greater than an input-dependent LAI value are valid

solutions. More information must be provided to the algorithm in such cases to localize

the solution (Zhang et al., 2000). Fourth, a unique solution cannot be expected in the

operational case of input uncertainties and algorithm imperfections. Thus, one can at best

derive a distribution of possible solutions and characterize this distribution by its mean

and variance. This is illustrated in Fig. 2 where the input plus its uncertainty are shown as

an ellipse (panel a). Every point in this ellipse is a valid input. The algorithm evaluates all

corresponding solutions, that is, the solution distribution function (panel b). The

12

dispersion of this distribution is an index of retrieval quality, and is in general larger than

input uncertainties.

3.3 Energy Conservation as a Constraint

The number of valid solutions may be unacceptably large in view of simplifying

assumptions in the algorithm and errors in input data. Therefore, the constraint provided

by the law of energy conservation on the inverse problem is valuable in obtaining

meaningfully localized solutions (Knyazikhin et al., 1998b). This principle is utilized in

the MODIS LAI/FPAR algorithm as follows. The model-based LUT entries are BRFs

parameterized in terms of three basic components of the energy conservation law,

namely, canopy transmittance and absorptance whose spectral variation can be explicitly

expressed via the leaf spectrum and two canopy structure specific wavelength

independent variables. This facilitates comparison of spectral values of BRFs with

spectral properties of individual leaves, which is a rather stable characteristic of a green

leaf. It allows the algorithm to admit only those LAI values for which the modeled BRFs

agree with the energy conservation law at any wavelength of the solar spectrum, thus

allowing a significant reduction in the number of retrieved solutions. Extensive

prototyping of the algorithms with data from different sensors (AVHRR, Landsat,

SeaWiFS) indicates that this constraint significantly enhances the accuracy of LAI/FPAR

retrievals (Tian et al., 2000; Wang et al., 2001).

13

3.4 Spectral Invariance

The extinction coefficient in vegetation canopies was treated by Ross (1981) as

wavelength independent in view of the size of the scattering elements (leaves, branches,

twigs, etc.) relative to the wavelength of solar radiation. This spectral invariance results

in a relation between canopy transmittance, t(λ0), and absorptance a(λ0) at a reference

wavelength λ0 to transmittances t(λ) and absorptances a(λ) at all other wavelengths λ in

the solar spectrum (Knyazikhin et al., 1998a),

)()(1

)(1)( 0

0 λλωλω

λ ttt

t

p

p

−−

= , )()(1

)(1

)(1

)(1)( 0

0

0 λλωλω

λωλωλ aa

−−

−−

=a

a

pp

(1)

where ω is the sum of leaf hemispherical reflectance and transmittance; pt and pa are

canopy structure dependent variables (therefore wavelength independent, but spatial

resolution dependent). The importance of these relations is two-fold. The size of the LUT

is independent of the number of spectrally dependent inputs ingested by the algorithm

since wavelength dependencies can be resolved from reference wavelength entries and

knowledge of pt and pa. Second, the scale dependence of the LUT, because of pt and pa,

facilitates validation of coarse scale retrievals with fine scale field measurements, as

discussed in Tian et al. 2000.

14

4 PERFORMANCE OF THE MODIS LAI/FPAR ALGORITHM

This section presents results that illustrate the performance of the main Look-Up-Table

(LUT) algorithm with respect to input and model uncertainties, dependence on spatial

resolution, impact of biome misclassifications, reflectance saturation, etc.

4.1 Input and Model Uncertainties

The uncertainties in input surface reflectances and models used to generate the LUT set a

limit to the quality of retrievals and their specification is critical to production of LAI and

FPAR fields of maximum possible quality. The in-orbit radiances measured by space-

borne sensors require corrections for calibration and atmospheric effects and this

introduces uncertainty in the surface reflectance products. For this reason, the use of only

the two bands with the smallest uncertainties, red and NIR (10% and 20% respectively),

as inputs to the functional algorithm is likely to be a permanent situation.

Model uncertainty characterizes the ability of models to approximate natural

variability, which in general can be quite high. These uncertainties depend on the

temporal and spatial resolution of the data as well. Ignoring model uncertainties in a

retrieval algorithm leads to destabilization of the convergence process because an

increase in input accuracy does not lead to better localization of the solutions (Wang et

al., 2001).

15

The convergence property of an algorithm guarantees increasingly accurate retrievals

with increasingly accurate inputs. To assure convergence, both input and model

uncertainties should be known. In general, this is not the case, especially with respect to

model uncertainties. Convergence can be assured by introducing a stabilization parameter

as described in Wang et al. (2001) which allows a correct specification of the overall

uncertainty in the inverse problem. The quality of retrieval is a function of the overall

uncertainty, denoted here as ε.

A measure, termed the retrieval index (RI), is defined as the number of pixels for

which the algorithm reports a solution relative to the total number of attempted pixels.

This variable does not characterize the quality of LAI and FPAR fields, but shows the

global coverage of the retrieved LAI and FPAR fields. The RI increases with an increase

in ε, but the reliability of retrievals decreases (Fig. 3). If ε underestimates the true overall

uncertainty, the algorithm fails to localize a solution, thus resulting in low values of RI.

On the other hand, if ε is an overestimate, the algorithm admits a large number of

solutions, including non-physical solutions, thus resulting in high values of RI. A critical

value of ε is one that optimally approximates the unknown true overall uncertainty. This

is estimated as the value of ε for which 95% of all land pixels for which the algorithm

fails were non-vegetated areas or regions where the data were corrupted due clouds or

atmospheric effect (Tian et al., 2000; Wang et al., 2001).

16

Figure 3. Dependency of the Retrieval Index (RI) on relative uncertainties (epsilon) in measurements and model simulations. The red and near-infrared reflectances input to the algorithm were from July 1989 LASUR (atmospherically corrected LAnd SUrface Reflectances in the red and near-infrared channels of the Advanced Very High Resolution Radiometer (AVHRR) at global scale, 1/7 degree resolution; one week temporal resolution for 1989) data set (Berthelot et al., 1997). From Tian et al. (2000).

4.2 NDVI and LAI/FPAR Relations

The non-linear relation between NDVI and LAI and the near-linear relation between

NDVI and FPAR reported in several studies has a physical basis as described in Myneni

et al. (1995) and Knyazikhin et al. (1998a). Therefore, the relation between NDVI,

evaluated with red and near-infrared reflectances input into the algorithm, and the

retrieved LAI/FPAR values, used to test the physics of the algorithm, will be examined.

Figure 4 shows the distributions of the retrieved LAI and FPAR values with respect to

the NDVI in the case of broadleaf forests. Relations between these retrieved LAI/FPAR

values and measured NDVI conform to both theoretical and empirical expectations. The

scatter in the relations highlights the merit of the algorithm versus a direct estimation

from NDVI. Panels c and d of Fig. 4 show scatter plots of input reflectance data from

17

successful and unsuccessful retrievals in the red and near-infrared space. This distribution

provides insight on where and why the algorithm failed.

Figure 4. NDVI vs. LAI / FPAR relations. The scatter plots show, for broadleaf forests, (a) the LAI-NDVI relationship, (b) NDVI-FPAR relationship, (c) retrieved and (d) non-retrieved pixels in RED-NIR space. The input reflectance data are July 1989 LASUR data set (Berthelot et al., 1997). From Tian et al. (2000)

For successful retrievals, the surface reflectances range from about 0.02-0.16 in the

red band and 0.10-0.42 in the near-infrared band. The algorithm tends to fail when (a) the

red reflectance is less than 0.03, i.e., the NDVI is very large, (b) red and near-infrared

reflectances both large, i.e., pixels are near the soil line and the NDVI is small, and (c)

intermediate cases. If the red reflectance is too small, the uncertainty is large and the

18

probability of a retrieval decreases. When a pixel is near the soil line in reflectance space,

it is probably either a non-vegetated pixel or the data is corrupted by clouds, and the

algorithm correctly identifies such cases. To understand the behavior of the algorithm for

intermediate values of reflectances, consider the NDVI contour as shown in panel d. For

the same value of NDVI, some pixels result in a retrieval while the others not. That is, the

algorithm utilizes information on canopy spectral and structural properties, instead of

NDVI, especially when it ingests three, four or even seven spectral bands and multi-angle

data. Only when a pixel falls within the spectral and angular space specified in the LUT,

a value for LAI is retrieved. Else, the algorithm returns a failure, even if the NDVI is

reasonable. It is likely that the non-retrievals correspond to biome mixtures, whose

probability is larger at coarse resolutions.

4.3 Impact of Biome Misclassification

The assumption that vegetation within each pixel belongs to one of the six biomes

impacts the LAI/FPAR retrievals. Results shown in Table 4 were obtained by running the

algorithm six times per pixel, each time with a different biome's LUT. This simulates the

effects of biome misclassification on the retrievals. With misclassification, either the RI

is low and/or the retrieved LAI values are incorrect. In case of misclassification between

distinct biomes, the results are predictable On the other hand, when misclassification

happens between spectrally and structurally similar biomes, perhaps because of coarse

spatial resolution, the impact on LAI/FPAR retrievals is difficult to assess.

19

Misclassified Biome Type

Grasses and

Cereal Crops

Shrubs Broadleaf

Crops Savannas

Broadleaf Forests

Needle Forests

Grasses and Cereal

Crops

91.53 1.20

88.54 1.25

89.60 1.40

88.68 1.36

27.63 1.29

29.00 2.01

Shrubs 87.67 1.03

92.66 1.41

91.53 1.54

91.73 1.51

47.34 1.51

46.37 1.99

Broadleaf Crops

87.93 1.85

70.33 1.83

74.03 2.10

71.29 2.04

14.80 2.42

19.52 3.71

Savannas 78.02 1.51

79.91 2.08

80.25 2.29

79.65 2.25

41.31 2.22

44.33 2.95

Broadleaf Forests

55.02 1.92

63.23 3.30

61.40 3.44

61.32 3.45

39.30 4.01

33.59 4.65

BCM Biome Type

Needle Forests

76.75 1.64

85.74 2.92

84.92 3.21

84.78 3.18

46.38 2.98

54.54 4.00

Table 4. Effect of biome misclassification on LAI retrievals. Data are from the July 1989 LASUR data set (Berthelot et al., 1997). The two entries in each box are the Retrieval Index and the mean LAI values, respectively. From Tian et al. (2000).

4.4 Scale Dependence of the Algorithm

The reflectance of a vegetation canopy is scale dependent. With a decrease in spatial

resolution of satellite data, the pixels are likely to contain an increasing amount of

radiative contribution from the background (Tian et al., 2000). This manifests as changes

in the location of reflectance data in the spectral space with changing spatial resolution.

Understanding the relation between such changes and LAI/FPAR variations with

resolution is key to accomplishing the scaling that is required in the validation of large

area retrievals with point field measurements. The MODIS LAI/FPAR algorithm

addresses this issue explicitly through structure dependent parameters pt and pa,

20

introduced earlier (Section 3.3), which imbue scale dependence to the algorithm via

modifications to the LUTs (Tian et al., 2000).

The scale dependence of the algorithm is illustrated here with retrievals from

algorithm runs on Landsat data with both Landsat (30m) and LASUR LUTs, and on

LASUR data with Landsat and LASUR LUTs (Table 5). When Landsat data and Landsat

LUT are used, the retrieved LAI values vary from 0-2.5 for grasses, 5-7 for broadleaf

forests and 1.5-6 for needle forests. The same runs with LASUR LUT result in unrealistic

retrievals - large LAI values and/or low retrieval indices because the location of fine

resolution data even in low LAI canopies map to locations of coarse resolution

reflectance data of dense canopies. Likewise, algorithm runs on LASUR data with

Landsat LUT result in low values of LAI for all biomes.

Coarse resolution reflectance data have larger radiative contributions from the

background in comparison to fine resolution data. While the LAI of the imaged scenes

may be identical, the location of reflectance data in the spectral space changes with

resolution. In the design of the MODIS LAI/FPAR algorithm, the three-dimensional

radiative transfer problem is split into two sub-problems. The first, the black-soil

problem, describes a vegetation radiation regime for the case of a completely absorbing

background beneath the canopy. The second, the S-problem, describes the radiation

regime due to interactions between the underlying surface and the canopy that includes

contributions from the background. At finer resolutions the contribution of the S-problem

is negligible, especially at high LAI values. For example, the retrieval indices from

Landsat data and Landsat LUT runs can be as high as 51% (broadleaf forests) using just

21

the black-soil problem compared to 54% when the S-problem is included. With coarse

resolution LASUR data, however, only using the black-soil problem lowers the RI from

39% to 31%, thus highlighting the importance of background contributions.

LASUR DATA LASUR LUT LANDSAT LUT

Biome Type Retrieval Index Mean LAI Retrieval Index Mean LAI Grasses &

Cereal Crops 91.53 1.20 91.6 1.07

Shrubs 92.66 1.41 96.4 0.92 Broadleaf

Crops 74.03 2.09 80.1 1.17

Savannas 79.65 2.25 85.4 1.61 Broadleaf Forests

39.30 4.01 41.8 2.62

Needle Forests 54.54 3.99 41.8 1.66 LANDSAT DATA

LANDSAT LUT LASUR LUT Biome Type Retrieval Index Mean LAI Retrieval Index Mean LAI Grasses and Cereal Crops

90.7 1.87 87.5 3.62

Broadleaf Forests

53.9 5.79 39.2 6.21

Needle Forests 57.9 4.11 4.7 3.39 Table 5. Retrieval indices and mean LAI values from algorithm runs on Landsat and LASUR data with Landsat and LASUR LUTs when the S-problem is included. From Tian et al. (2000).

4.5 Reflectance Saturation

In the case of dense canopies, the reflectances saturate and are therefore insensitive to

changes in LAI (Fig. 5). The canopy reflectances are then said to belong to the saturation

domain (Knyazikhin et al., 1998b). The reliability of parameters retrieved under

conditions of saturation is low, that is, the dispersion of the solution distribution is large

22

(Fig. 2, case 5). The frequency of LAI retrievals under saturation also increases with

increasing uncertainties. The saturation domain can be avoided if more information can

be provided to the algorithm in the form of multi-angle and multi-spectral data (Zhang et

al., 2000). Below, the saturation domain problem with multi-angle retrievals is illustrated.

Figure 5. Illustration of the convergence property of the LAI model from different base soil conditions. From Shabanov et al. (2000).

Example algorithm runs with multi-spectral multi-angle data over Africa from the

POLDER instrument (Leroy et al., 1997), assuming a mean overall uncertainty of 20%,

were performed. The LAI saturation frequency decreases with an increase in the number

of view angles (Table 6). This is evidence of the enhanced information content of multi-

angle data, as it helps localize a value of LAI. As expected, the saturation domain is

rarely encountered in sparse biomes, such as grasses and shrubs.

Saturation domain retrievals are flagged in the quality assessment (QA) files

accompanying the MODIS LAI/FPAR product. For each such retrieval, the lower bound

23

of the LAI domain can be evaluated. This value is to be interpreted as follows: all values

of LAI above this threshold, up to the maximum value of 7.0, are valid solutions (see

case 5 in Fig. 2a). The algorithm reports a value equal to the arithmetic mean of these

solutions.

BCM Biome Type

(å = 0.2) (%)

# of View Angles

Grasses & Cereal Crops

Shrubs Broadleaf

Crops Savannas

Broadleaf Forests

Retrieval Index 1 6

99.5 84.0

99.90 80.8

99.2 73.4

98.1 74.9

40.6 18.9

LAI Saturation Frequency

1 6

0.01 0.0

0.02 0.01

6.26 2.52

9.76 6.18

30.1 17.4

Table 6. Retrieval Index and frequency of LAI under saturation for POLDER data (1-16 November, 1996) over Africa with mean overall uncertainty of 0.2. From Zhang et al. (2000).

24

5 THE MODIS LAI / FPAR PRODUCT

The MODIS LAI/FPAR product is produced at 1km spatial resolution daily (MOD15A1)

and composited over an 8-day period based on the maximum FPAR value. The 8-day

product (MOD15A2) is distributed to the public from the EROS Data Center Distributed

Active Archive Center (EDC DAAC). Monthly (MOD15A3) global coverage layers will

be produced at Boston University. The products are projected on the Integerized

Sinusoidal (IS) 10-degree grid. The globe is tiled, occupying 452 of them, for production

and distribution purposes into 36 tiles along the east-west axis, and 18 tiles along the

north-south axis, each representing 1200x1200km.

The product files contain four scientific data sets, output as 2 dimensional HDF EOS

grid fields of 1200 lines by 1200 samples. All fields are produced using the HDF uint8

data type, which is an unsigned 8 bit integer variable whose values may range from 0 to

255. The values are stored in their digital form with a scale-factor (gain) and offset which

is applied to transform the stored values to their biophysical counterparts for analysis.

The quality control variables are bit-flags and integer measures without a gain or offset

(Table 7). The product files also contain a considerable amount of extra information that

describes various properties of the data. The majority of this information is classic

metadata, describing the geolocation, quality, and source of the tile and pixel data.

Visual inspection of the MOD15A2 product is a quick way to check on the quality of

the data. A proper interpretation of a data layer also requires consulting the

corresponding QA information. Because the algorithm is more sensitive to cloud

25

contamination than the cloud masking decision threshold used in the cloud mask

algorithm, a set of multiple 8-day period coverages must be used to build an accurate

global picture.

SDS Variable Data Type Fill Value Gain Offset Valid Range Fpar_1km Uint8 255 0.01 0.0 0, 100 Lai_1km Uint8 255 0.10 0.0 0, 100 FparLai_QC Uint8 255 N/A 0.0. 0, 254 FparExtra_QC Uint8 255 N/A 0.0 0, 254

Table 7. Summary of the MODIS LAI/FPAR Scientific Data Sets (SDS). The expression used to decode the Digital Numbers to the corresponding analytical value is: analytical value = scale factor x (digital number – offset). From Knyazikhin et al. (1999).

To keep track of coverage for each 8-day set, a visualization tool was developed to

render this information (Fig. 6). The MODLAND Integerized Sinusoidal Grid and a list

of the files acquired are read by the tool. For each file in the list, the tool uses a flood-fill

algorithm and a pre-generated set of points to color the corresponding tile. In the process

of coloring the map, the tool tallies the total number of tiles and those pre-defined to

consist primarily of land pixels (Fig. 7).

Outside of two short interruptions, continuous temporal coverage of the product is

publicly available from the EOS Data Gateway from as early as June 9, 2000. Full

coverage of North America and the Australian continent was acquired by the second A2

period. South America, Africa and Europe were fully covered in the third A2 period.

Tiles in the Northern Siberia portion of the Asian continent were the last holdout, not

being produced till the October 7, 2000, A2 period. Since the September 13, 2000, A2

26

period, regular coverage has been near global. Additionally, tiles containing groups of

small islands are also regularly produced.

Figure 6. 8-Day Tiles produced during year 2000. Global tile coverages of MOD15A2 data available from 06/09/00 (top left), down the columns, to 12/25/00 (bottom right).

27

A2 Tiles Produced During 2000

0

50

100

150

200

250

300

350

161

177

201

241

257

273

289

305

321

337

353

Julian Date

Nu

mb

er o

f T

iles

Non-land tiles

Land tiles

Figure 7. Number of total and land tiles produced for each 8-day period. Non-land tiles are those tiles with either a small fraction of continental land pixels or only a group of islands.

With regular near-global coverage it becomes possible to create monthly or seasonal

(Fig. 8) composited data layers. Producing the monthly A3 layer is handled in the same

way as the creation of an A2 layer from 8 A1 daily coverages. For each pixel, the

maximum FPAR value is found from the input data sets, the corresponding LAI value

from that day is then used, whether or not that is the maximum value for that time period.

While useful for scientific analysis, it is hard to extract useful information through

visual inspection. Limited by the size of monitors, it is impossible to view an entire layer

that measures 43,200 x 21,600 pixels. Also of practical consideration is the fact that a

single global layer is approximately 1 gigabyte in size, which is larger than a number of

programs are capable of handling.

28

Figure 8. Color-coded maps of MODIS LAI and FPAR fields for the boreal summer of year 2000. Each pixel in the image is the average of 30 x 30 MODIS pixels (water pixels not included). The maximum FPAR value from the set of MOD15A2 data was chosen for each pixel for the FPAR image. The LAI value from the same date as the maximum FPAR value was chosen for the LAI image.

To allow for both rigorous and casual observation, a spatially averaged image is

produced. In these images, each pixel represents a 30x30 km area. Water pixels are

subtracted from the total number of pixels when forming the average of new land pixels.

New pixels with a majority of water pixels as constituents are designated water pixels. As

can be seen by a comparison of pre- and post-sampled equally sized pictures of Australia

in Fig. 9.

29

Figure 9. Comparison of pre- and post-spatially composited FPAR images of Australia. Both images are equivalent density slices (same as the one used in the Fig. 7 FPAR image), but different color tables. Each pixel represents (a) 1x1 km2 and (b) 30x30 km2.

Production of spatially coarsened images must do without the benefit of associated

QA information. For the northern hemisphere growing season composite layers, there

were enough input layers to guarantee that statistically anomalous low values were not

included in the averaging process. A general quality assessment for maximum composite

images can be given as a ratio between the number of pixels for which a main algorithm

value existed and the total number of A2 land pixels used to generate A3 land pixels.

Within the MODIS team, considerable attention has been paid to implement a set of

quality control protocols that help users match data sets to their applications. Quality

control measures are produced at both the file (10° tile level) and at the pixel level. At the

tile level, these appear as a set of EOSDIS core system (ECS) metadata fields. Quality

control information, at the pixel level, is represented by one or more separate data layers

in the HDF EOS file. Quality-scoring schemes vary by product at the pixel level.

The FPAR / LAI quality control variable summary (Table 8) includes a description of

the meaning of each bit in the QA field. Proper interpretation of the bits is given in the

rightmost column. It is important to note that if any of the cloud cover bits are set, the

level of data quality will be decremented by one.

30

Bit-field Binary values Description Interpretation

MODLAND

00 01 10 11

Highest quality Good quality Not produced, cloud Not able to produce

Saturation condition

ALGORITHM PATH

0 1

Empirical method Main R-T method

Back-up Algorithm Main algorithm

DEAD DETECTOR 0 1

No dead detectors Dead detector

CLOUDSTATE

00 01 10 11

Cloud free Cloud Covered Mixed clouds Not set, assume clear

Marked cloud free Cloudy Cloudy

SCF_QC

00 01 10 11

Best Model Result Good quality Use with caution Could not retrieve

Table 8. The FPAR / LAI quality control variable summary chart. Descriptions of the meaning of the bits found in the FparLai_QC variable. Note that the Dead detector record was first inserted into the bit-field with the A2 period starting on September 13, 2000.

In the analysis of QA information associated with LAI and FPAR values, the different

QA field permutations are turned into 4 main categories. Cloud-free pixels can be labeled

as being produced via the main algorithm, main algorithm under conditions of saturation

or the back-up algorithm. Pixels marked as suffering from cloud contamination, if not

given a fill value, primarily generate their values from the back-up algorithm.

Even though the cloud-state bits might be set to indicate clear conditions, there still

might be signal-degrading clouds in that pixel. This is made clear through visual

inspection of the LAI and QA fields from two A2 periods of the tile containing the

Amazonia region (Fig. 10). The center region of the tile from July 3, 2000, shows

scattered clouds, but the pixels with low LAI values are contiguous over that area. The

same area of the tile from the next A2 period is shown to have had its values produced

31

via the main algorithm. The difference between the small early values and the large latter

values in the same area can be attributed to clouds or other atmospheric effects.

To verify this hypothesis a data set consisting of 29 pairs of consecutively dated A1

tiles were used. A difference was calculated for every pixel where the two days were

marked cloud-free, one day was produced with the main LUT algorithm and the other

produced through the back-up NDVI based algorithm. Plotting these differences against

the value of the main algorithm shows that as they increase, the back-up values

increasingly lag behind (Fig. 11). At small values, normal distribution of the inherent

error in the data is disrupted by the fact that values may not be less than zero, therefore

producing negative differences. Negative FPAR differences can also result from its

insensitivity to NDVI changes for small NDVIs (cf. Fig. 4.b.). Leaf Area Index, on the

other hand, is highly sensitive to NDVI variations at the lower range of values, which can

allow for large differences.

The A1 daily product shows a high correlation (Table 9) between values produced by

the main algorithm and the increasing difference of those produced by the back-up

algorithm. Little variation in physical condition of land cover can be expected on a 1x1

km scale over a 24-hour period. Therefore, the causes of variations in values will be due

to a bias between the algorithms and uncertainties in the model and inputs. For all these

curves, LAI main – LAI back-up ≈ LAI main (i.e., LAI back-up ≈ 0). This means that the main

algorithm fails when the pixels are corrupted due to clouds or atmospheric effects.

32

Figure 10. Eight-day composite MODIS LAI product for a region in the Amazonia for two periods in July 2000. The corresponding quality assessment (QA) fields are shown in the two panels below. The main algorithm is the Look-Up-Table based method and the back-up algorithm is the NDVI based method. The latter is used when the main algorithm fails. Note the low LAI values in the tile center during the early period due to clouds.

33

Biome 1 Biome 2 Biome 3 Biome 4 Biome 5 Biome 6 FPAR 0.798 0.966 0.914 0.882 0.904 0.882 LAI 0.952 0.995 0.943 0.974 0.955 0.873

Table 9. Correlation between average (main – back-up) differences and gradated values of LAI and FPAR produced by the main algorithm for each difference. Pearson correlation coefficients, P < 0.0001.

A more detailed examination of the data reveals that the majority of back-up values

are zero. For all biomes, the number of pixels with a value of zero was an order of

magnitude larger than the number of pixels distributed to any other value. A simple

regression analysis comparing all the non-zero back-up values with their corresponding

main algorithm values shows little agreement (Table 10).

Biome 1 2 3 4 5 6

R-Squared 0.5202 0.2375 0.0574 0.0494 0.1730 0.0077 Slope 0.6694 0.7556 0.2482 0.2348 0.4091 -0.1259

Table 10. Regression results between main and back-up values for corresponding pixels from consecutively dated A1 data.

Further proof of the greater sensitivity of the main algorithm to clouds is found by

charting resulting QA values as a function of latitude (Fig. 12). Each of the six biomes

shows a peak in main algorithm failure at tropical latitudes. The same latitude band is the

site of frequent cloud cover due to the Inter-Tropical Convergence Zone. Conversely, a

higher percentage of main algorithm pixels are produced in mid-latitude high-pressure

bands.

Consistency of the product can be gauged by examining the distribution values, the

distribution of differences between values, and by identifying anomalous biases. LAI and

34

FPAR values generated by the main algorithm appear to be similar under both regular

and saturation conditions (Fig. 13). There also appear to be similarities between the

distributions of back-up LAI and back-up FPAR.

Ranges between main and back-up algorithms are, overall, smaller for LAI than for

FPAR (Fig 13). With the exception of the two dense biomes, broadleaf and needle

forests, 70% of LAI differences fall within the 20% uncertainty range. Outcomes for

FPAR are worse, with only 20 – 40% of differences lying within the same range, and

again, the forest biomes show the least coordination.

Although overall FPAR differences are larger than those of LAI, their magnitudes are

proportional (Fig. 14). Small FPAR differences are more likely to occur with small LAI

differences (Fig. 15). Low LAI differences may still be related to higher FPAR variations

because of the non-linearity of the LAI-NDVI curve as compared to the NDVI-FPAR

curve (cf. Fig 4).

35

Mean Algorithm Differences

-6.0

-4.0

-2.0

0.0

2.0

4.0

6.0

8.0

0.3 1.0 2.0 3.0 4.0 5.0 6.0 LAI main

∆ LA

I

Biome 1

Biome 3

Biome 4

Biome 5

Mean Algorithm Differences

-1.0

-0.5

0.0

0.5

1.0

0.03 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 FPAR main

FP

AR

Biome 1

Biome 3

Biome 4

Biome 5

Mean Algorithm Differences

-4.0

-2.0

0.0

2.0

4.0

6.0

0.25 1.00 2.00 3.00 4.00 5.00 6.00 LAI main

∆ LA

I

Biome 1

Biome 2

Biome 3

Biome 4

Biome 5

Biome 6

Mean Algorithm Differences

-0.8-0.6

-0.4-0.2

00.2

0.40.6

0.8

0.03 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 FPAR main

∆ F

PA

R

Biome 1

Biome 2

Biome 3

Biome 4

Biome 5

Biome 6

Figure 11. Differences between values produced by the main and by the back-up algorithm as a function of values from the main algorithm. Data derived from MOD15A1 tiles (top 2) for 6 pairs of days for 2 tiles from central Africa and (bottom 2) a pair of days using 17 tiles over the North and South American continent.

36

Q A R e s u lts by Lat i tude, B i o m e 1

0 .0

0 .2

0 .4

0 .6

0 .8

1 .0

Lat i tude

M ain

Sa tu ra t i on

B a c k u p

Q A R e s u lts by Lat i tude, B i o m e 4

0 .0

0 .2

0 .4

0 .6

0 .8

1 .0

Lat i tude

M ain

Sa tu ra t i on

B a c k u p

Q A R e s u lts by Lat i tude, B i o m e 2

0 .0

0 .2

0 .4

0 .6

0 .8

1 .0

Lat i tude

M ain

Sa tu ra t i on

B a c k u p

Q A R e s u lts by Lat i tude, B i o m e 5

0 .0

0 .2

0 .4

0 .6

0 .8

1 .0

Lat i tude

M ain

Sa tu ra t i on

B a c k u p

Q A R e s u lts by Lat i tude, B i o m e 3

0 .0

0 .2

0 .4

0 .6

0 .8

1 .0

Lat i tude

M ain

Sa tu ra t i on

B a c k u p

Q A R e s u lts by Lat i tude, B i o m e 6

0 .0

0 .2

0 .4

0 .6

0 .8

1 .0

Lat i tude

M ain

Sa tu ra t i on

B a c k u p

Figure 12. Average QA value of pixels as a function of latitude. Each chart is an average of the values from four A2 global coverage periods, whose initial dates are 6/25, 7/19, 8/20 and 8/28. For each degree of latitude all pixels in each QA condition, that were marked as ‘not cloudy, are divided by the total number of non-cloudy pixels for that biome at that latitude.

37

Distribution of Main LAI Values

0%

20%

40%

60%

80%

100%

0.1

1.0

1.9

2.8

3.7

4.6

5.5

6.4 LAI

Pix

el %

, by

Bio

me

Biome 6

Biome 5

Biome 4

Biome 3

Biome 2

Biome 1

Distribution of Main FPAR Values

0%

20%

40%

60%

80%

100%

0.02

0.18

0.34

0.50

0.66

0.82 FPAR

Pix

el %

, by

Bio

me

Biome 6

Biome 5

Biome 4

Biome 3

Biome 2

Biome 1

Distribution of Saturation LAI Values

0%

20%

40%

60%

80%

100%

0.1

1.0

1.9

2.8

3.7

4.6

5.5

6.4 LAI

Pix

el %

, by

Bio

me

Biome 6

Biome 5

Biome 4

Biome 3

Biome 2

Biome 1

Distribution of Saturation FPAR Values

0%

20%

40%

60%

80%

100%

0.02

0.18

0.34

0.50

0.66

0.82 FPAR

Pix

el %

, by

Bio

me

Biome 6

Biome 5

Biome 4

Biome 3

Biome 2

Biome 1

Distribution of Backup LAI Values

0%

20%

40%

60%

80%

100%

0.1

1.0

1.9

2.8

3.7

4.6

5.5

6.4 LAI

Pix

els,

by

Bio

me

Biome 6

Biome 5

Biome 4

Biome 3

Biome 2

Biome 1

Distribution of Backup FPAR Values

0%

20%

40%

60%

80%

100%

0.02

0.18

0.34

0.50

0.66

0.82 FPAR

Pix

el %

, by

Bio

me

Biome 6

Biome 5

Biome 4

Biome 3

Biome 2

Biome 1

Figure 13. Distribution of LAI and FPAR values from the global coverage 8-day period 06/25 to 07/02. Graph values are the percentage of pixels, for the given biome, with that value. These percentages are converted into the percentage contribution for the given biome for the particular data value.

38

Cumulative Distribution of Differences

0.00.10.20.30.40.50.60.70.80.91.0

0.15

0.75

1.35

1.95

2.55

3.15

3.75

4.35

4.95

5.55

6.15

6.70

Absolute Value of ∆ LAI

Fra

ctio

n of

Pix

els,

by

Bio

me

Biome 1

Biome 2

Biome 3

Biome 4

Biome 5

Biome 6

Cumulative Distribution of Differences

0.0

0.10.2

0.30.4

0.5

0.60.7

0.80.9

1.0

0.00

0.08

0.16

0.24

0.32

0.40

0.48

0.56

0.64

0.72

0.80

0.88

0.96

Absolute Value of ∆ FPAR

Fra

ctio

n of

Pix

els,

by

Bio

me

Biome 1

Biome 2

Biome 3

Biome 4

Biome 5

Biome 6

Figure 14. Cumulative distributions of differences between main and back-up values of pixels over consecutive days. Data represents 6 pairs of days for 2 tiles from central Africa and a pair of days using 17 tiles over the North and South American continent.

39

Figure 15. Comparison of magnitudes of LAI and FPAR differences between main and back-up values. Data are derived from 17 pairs of consecutively dated tiles from North and South America.

40

6 Validation Without a program of rapid validation for these products, the algorithm-classified pixels

would go out for use to researchers whose models could have different sensitivities to

misclassification. Such a problem could lead to conflicts of model outcomes. Subsequent

indecision by policy-makers who would use those model results would negate the

primary usefulness of the product.

The responsibility for validation of the MODIS LAI/FPAR product will be shared

between the algorithm developers and validation investigators selected in response to

NASA research announcements. However, for global validation, it is recognized that

greater resources and coordination are required than are currently recruited. Thus,

MODLAND is applying significant effort to developing community wide validation

protocols and encouraging participation by data collectors and product users through the

aegis of international bodies such as the working group on calibration and validation of

the Committee on Earth Observation Satellites (Privette et al., 1998; Morisette et al.,

2000). These activities bring together the resources of various international teams

working on validation of land surface products from moderate resolution sensors.

Global validation requires field data from a range of sites representing a logical subset

of the Earth's land covers. The initial focus of our validation activities is at the EOS land

validation core sites as these are high priority Terra and Landsat 7 land validation

activities and product generation targets. The sites typically have a history of in-situ and

remote observations and are expected to facilitate both validation and early EOS science.

Centralized web based archiving of ASTER, MISR, MODIS, Landsat 7 ETM+, and

41

IKONOS products in relatively easy-to-use formats are planned for these sites. These

core sites together with several investigator volunteered sites comprise the LAI-net, an

informally coordinated array of sites for possible validation of LAI/FPAR products from

MODIS and other moderate resolution sensors (Table 11).

Name Country Biome EOS Core Site USDA BARC, MD U.S. Broadleaf cropland X Bondville, IL U.S. Broadleaf cropland X Gainesville, FL U.S. Broadleaf cropland Tapajos Brazil Broadleaf Forest X Hawaii U.S. Broadleaf Forest Harvard Forest, MA U.S. Broadleaf Forest X Park Falls, WI U.S. Broadleaf Forest X Uradry Australia Grassland X Osage, OK U.S. Grassland Konza, KS U.S. Grassland X East Anglia England Grassland X Vernon, TX U.S. Grassland BOREAS NSA Canada Needle forest X Cascades, OR U.S. Needle forest X EMATREF France Needle forest Yaqui Valley Mexico Shrub San Pedro Basin/ SALSA, AZ

U.S. Shrub X

Skukuza South Africa Shrub/woodland X New Zealand network

New Zealand Various

Canada Network Canada Various Mongu Zambia Woodland X Cerrado Brazil Woodland Safari-2000 Botswana Various Ruokolahti Finland Needle forest Kejimkujik Canada Broadleaf forest

Table 11. LAI-net sites for validation of LAI/FPAR products from MODIS and other moderate resolution sensors.

42

6.1 Bigfoot

The BigFoot project grew from a workshop that was attended, primarily, by scientists

from the Long Term Ecological Research (LTER) Network. The purpose was to explore

validation protocols and scaling issues that would lead to an improved understanding of

MODIS land cover, LAI, FPAR and NPP products. The BigFoot validation program will

be working at four field sites. Biomes included in these sites are: boreal forest, tall-grass

prairie, mixed deciduous-conifer forest and a mixed corn and soybean agricultural system

(BOREAS NSA, Canada; Harvard Forest, MA.; Bondville, IL.; Konza, KS, respectively).

Intended products from these field campaigns are LAI, FPAR, NPP, and related

variables. Five-kilometer grids will be covered at each site with a 25m spatial resolution.

As MODIS has a resolution from 250 – 1000m, and BigFoot has a 25m resolution,

scaling effects must be looked at. In recent years there have been suggestions by

researchers that a separate science of scale should be developed, the main focus of which

would be the Modifiable Areal Unit Problem (MAUP), as first stated by Woodcock and

Strahler (1987). They discerned that a choice of scale depends on the output information

desired about the ground scene, methods used to extract information from the imagery,

and the spatial structure of the scene. Variation in classification outcomes can be due to

aggregating pixels into larger units (scale), and from making alternate combinations of

areal units of similar scale (aggregation).

Work on derivation of proper rules for scaling continues, but even without an

algorithm to make a choice by, fields affected by MAUP must still press on. To address

43

this, the science teams will explore the errors and information losses that accrue during

the extrapolation of field data to coarse grain (1km) surfaces (Campbell et. al 1999). It

will be attempted to determine if there is a fundamental grain size at each site, above

which error rates accelerate when modeling land cover, LAI, FPAR and NPP. The fine-

grained surfaces will be aggregated to several resolutions, up to 1km, to determine if

there exists such a fundamental grain size.

Through these analyses, it is hoped that error due to scaling differences versus error

due to algorithm definitions will be characterized. It is theoretically possible for no single

MODIS grid cell to accurately estimate land cover, LAI, FPAR or NPP, but for multiple

cell estimates within and across sites to be accurate. A cross-site comparison of

MODLAND and BigFoot surfaces will permit us to assess MODLAND data product

accuracy and gain an understanding of the source of errors at both site and cross-site

levels.

6.2 SAFARI 2000 Wet Season Campaign

Members of the Boston University Climate and Vegetation Group participated in the

SAFARI 2000 wet season field campaign in Southern Africa from March 3-18, 2000.

Ground measurements of LAI, FPAR, leaf and canopy hemispherical reflectance and

transmittance, and directional canopy reflectance were made using the LAI-2000 plant

canopy analyzer, AccuPAR ceptometer, LI-1800 portable spectroradiometer and ASD

handheld spectroradiometer (no endorsements intended here and through out). LAI and

FPAR were intensively measured at 4 different sites; Pandamatenga, Maun, Okwa and

44

Tshane (from north to south in Botswana), where the vegetation ranged from moist

closed woodland to arid sparsely-shrub covered grassland.

At each of the four sites, measurements were collected every 25m along three parallel

750m long East-West transects. These transects were 250m apart along the North-South

axis. In addition, data were also collected on a 250m by 300m grid at every 50m. The

measured LAI distributions at the four sites are shown in Fig. 16 together with the

distribution assembled from the 1km MODIS LAI product over pixels in the vicinity with

the same land covers. The agreement between the two is noteworthy considering high

variability in both field data and the product.

Figure 16. Comparison of MODIS LAI product with SAFARI 2000 wet season field campaign measurements. Figures show the distribution of LAI values derived from field measurements and evaluated with the MODIS LAI/FPAR algorithm. Dispersions derived from field measurements are higher than those derived from MODIS retrievals because the MODIS LAI/FPAR algorithm only accounts for the most probable situations encountered in reality. Mean MODIS LAIs agree with mean LAI values derived from field measurements. From Myneni et al. (2001).

The scaling problem in the validation of moderate resolution products with higher

resolution satellite imagery and field measurements can be stated as follows. Scaling is

defined as the process by which it is established that the LAI/FPAR values derived from

45

coarse resolution sensor data are the arithmetic averages of LAI/FPAR values derived

independently from fine resolution sensor data over the same region (Tian et al., 2001).

The variables pt and pa, which imbue scale dependence to the algorithm via modifications

to the LUTs, can be derived from model calculations and measurements of leaf and

canopy spectral properties (Panferov et al., 2001; Tian et al., 2001). Ground based

measurements that allow specification of pt and pa are included in a prioritized list of

measurements needed for validation of MODIS LAI/FPAR product. Thus the transect

and grid point measurements collected in the field can be used to obtain the scaling

parameters pt and pa at spatial resolutions of interest.

The LAI maps of a 5×5km region at one of the sites, Maun, derived from MODIS

(1km), ETM+ (30 m) and IKONOS (4m) data, are shown in Fig. 17. The MODIS

LAI/FPAR algorithm was used in all cases, but with Look-Up-Tables adjusted for

resolution of input reflectance data from each of these sensors. The retrieved distributions

are shown in Fig. 18 together with field measurements. The agreement between the

various distributions illustrates the validity of the scaling approach and the MODIS LAI

product for this site. In Table 12, the higher resolution IKONOS-based LAI retrievals are

compared with both transect and grid field measurements. Again, the agreement between

the retrievals and measurements is noteworthy.

Experiment IKONOS LAI Measured LAI Transect N 1.47 ± 0.73 1.77 ± 0.82 Transect A 1.69 ± 1.31 1.14 ± 0.82 Transect B 2.31 ± 1.57 1.62 ± 0.57 Grid 2.02 ± 1.54 1.21 ± 0.68

Table 12. High-resolution IKONOS-based LAI retrievals compared to transect and grid field measurements at Maun, Botswana. Transect and grid measurements are explained in the text. From Tian et al. (2000).

46

Figure 17. LAI maps of a 5 x 5km region at Maun in Botswana retrieved by the LAI/FPAR algorithm with MODIS, Landsat ETM+ and IKONOS data. These are the temporally closest available images to the dates of the field measurements. From Myneni et al. (2001).

6.3 Summer 2000 Field season

During the 2000 summer field season, members of the Boston University Climate and

Vegetation Group participated in three validation campaigns in North America and

Finland. Four different instruments were available for the collection of field data.

Analysis of field data is ongoing. Validation of field data is a multi-step operation,

involving: cleaning the field data, comparing field measurements to data from higher

resolution satellites (e.g. Landsat and IKONOS) and finally cross-validating MODIS

MODIS Retrievals, Apr. 3, 2000

ETM Retrievals, Apr. 3, 2000

IKONOS Retrievals, Mar. 30, 2000

47

reflectances with those from the higher resolution satellites. This work is, as of yet,

incomplete.

Figure 18. Distributions of LAI retrievals shown in Fig. 16 together with field measurements at Maun, Botswana. From Myneni et al. (2001).

6.3.1 Instruments

6.3.1.1 LAI-2000

An LAI-2000 plant canopy analyzer is composed of an LAI-2070 control unit and an

LAI-2050 sensor head. The control unit has connectors for two sensors so that both may

be automatically calibrated. The sensor heads project a nearly hemispheric image onto

five detectors arranged in concentric rings (approximately 0-13, 16-28, 32-43,47-58 and

61-74 degrees). Radiation of wavelengths above 490nm is rejected from analysis. Caps

blocking various sectors may be placed over the detector to prevent operator interference.

There is both a single and a double unit method of data collection available with the

LAI-2000. If two units are used, both sensors are calibrated with each other in an open

area with uniform sky conditions. One instrument remains on a stand in the open area to

48

serve as a reference for above the canopy radiation while the other instrument is used to

take actual field readings. If only a single instrument is available. The preferred method

of data collection is to take an open sky reading, acquire several measurements and then

take another open sky reading. Given consistent atmospheric conditions, changes in

irradiance due to the sun’s position can be determined from the pre- and post-

measurement sky readings.

The LAI-2000 measures the attenuation of diffuse sky radiation at five zenith angles

simultaneously. LAI measurements were made during the early morning and late

afternoon under overcast sky conditions. The sensor is to be held horizontally over the

ground at a minimum height of 1m, moving higher up to keep above the local under-

story. Every data point represents an average of three separate measurements made in

close spatial and temporal proximity.

6.3.1.2 AccuPAR Ceptometer

An AccuPAR Ceptometer is a single device for measuring photon fluxes (µmol*m-2*s-1)

in a range from 300 to 700nm. The sensor is a light measurement bar with an

approximate 180° field of view along its 1m length. Measurements are made by holding

the light at a height of about 1m or over the local under-story. It should be held

horizontally, perpendicular to the plane of the sun, upwards for incident and downward

for reflected measurements, and in a direction perpendicular to the direction of the sun.

Accurate measurements are not made if the sun is too obscured to cast shadows. Every

49

data point represents an average of three separate measurements made in close spatial and

temporal proximity.

6.3.1.3 ASD Spectroradiometer

The ASD measures radiances by taking a sample reading of a lambertian panel before

making a field measurement. Both of these measurements need to be made under the

same atmospheric conditions. Measurements are taken in the range of 400nm to 1100nm

at a 1nm resolution. Fore optics are available to limit the field of view of the detector.

The number of samples used for integrating a single measurement is user-defined.

6.3.1.4 LI-1800

The LI-1800 is a device for measuring the leaf optical properties, specifically leaf

spectral hemispherical reflectances and transmittances. Measurements are in the spectral

range of 400nm to 1100nm. Not a field device, leaf samples are brought in from the field

and placed in its goniometer assembly.

6.3.2 Field Campaigns

6.3.2.1 Harvard Forest

Investigated from 07/21/200 to 07/25/2000, the Harvard Forest is primarily a broadleaf

forest area in western Massachusetts. Field instruments used included two LAI-2000s,

one ASD, and two AccuPARs. Due to a malfunction with the LI-1800, leaf samples were

50

obtained and measured in early September. Myself, along with ten other members of the

B.U Climate and Vegetation Group performed three main experiments during that time.

6.3.2.1.1 Grid Experiment

A 225x225m grid was laid out on relatively flat terrain just outside the southeast

boundary of Harvard Forest. This grid was delimited by a matrix of 9 rows, labeled "1 -

9", and 9 columns, labeled "A - I", giving a total of 81 points. Each cell-center point was

25m from its neighbors. The columns were oriented to run from magnetic north (row 1)

to south (row 9), the rows ran perpendicular to the columns from West (column A) to

East (column I), so that the center point of the grid would be labeled "E-5". Distances

were checked using tape measures and directions checked using compasses. These points

were later verified with a GPS. The gird center is located at the UTM coordinates of zone

18, 732083 E, 4712906 N.

Data was accumulated with each of our three instruments over this grid system. ASD

and LAI-2000 measurements were taken for every point on the grid. LAI-2000 data was

collected for the entire grid over a period of two days. PAR data was also collected on

two separate days, but the entire grid was not covered. All points from columns "A" - "F"

and "G1" - "G6" were sampled on one day, and only all of columns "A" - "F" were

measured on the other day. All measurements were taken starting at point "A1", then

going to "A9", "B9" to "B1", "C1" to "C9" and so on. Both canopy transmittance and

under-story reflectance measurements were taken with the ASD and the AccuPAR. A

characterization of the dominant over-story species in the grid area is depicted in graph 2

6.3.2.1.2 Road Experiment

51

Measurements were taken along two of the roads going through Harvard Forest in an

effort to determine the disruption they cause. Starting at the western opening of each

road, an AccuPAR measurement was made every 1/10th mile, as measured by a car's

odometer. There were 10 data points along Locust Opening Road and 14 points along

Prospect Hill Road.

6.3.2.1.3 Short transect Experiment

Measurements were taken in a single straight line over a distance of approximately 330

feet. This line was traversed three times, with ten data points being acquired each time,

each approximately 33 feet from the previous point. Both canopy transmittance and

under-story reflectance measurements were taken. The transect was located in the area

surrounding the fire tower, so that accompanying directional canopy reflectance

measurements could be made with the ASD spectroradiometer. The dominant over-story

species in this area was oak.

6.3.2.2 Kejimkujik National Park

From 08/10/2000 to 08/18/2000, myself along with two members of the Canadian Center

for Remote Sensing acquired ground truth data in Kejimkujik (Keji) National Park. Keji

has over 300km2 of dense mixed forest cover representative of the larger surrounding

region. We were able to make measurements at 28 sites and along a 1.5 km transect, all

together covering a thin 14.5km swath along the eastern portion of the forest. Resources

for this campaign included one AccuPAR, one camera with a hemispherical lens, and

three LAI-2000s. Measurements were made using both the single- and double-LAI-2000

52

method. The internal model used by the LAI-2000 is geared towards gap fraction analysis

of broadleaf areas. The CCRS processing of needle forest areas used the equation,

L = (1 - α) L e γE / ΩE (2)

as given in Chen (1996) to derive the LAI of a plant canopy. Here, L denotes LAI; α is

the woody-to-total plant area ratio; L e is the effective LAI measured by the instrument; γE

is the needle-to-shoot area ratio; and ΩE is the foliage element clumping index.

In keeping with the requirements of the CCRS Landsat LAI map validation program,

measurements at all sites covered 30m. For each site, the ground was examined to

determine slope. At sites with moderate slopes, in-canopy measures with the LAI-2000s

were made along and across the slope. Measurement tracks were aligned perpendicular

to- and parallel to the solar plane at sites with no observable slope. Four measurements

were made along each track. The two tracks crossed each other at the sites’ center.

AccuPAR measurements were taken along a random direction at an average distance of

every 2m. A minimum distance of 100m separated individual sites from each- other.

Preliminary results from CCRS analysis (White et. al 2000) support their empirically

derived algorithm (Chen and Cihlar, 1996; Brown et al., 2000) for generating LAI from

the AVHRR.

6.3.2.3 Ruokolahti

Field time in the needle leaf forest at Ruokolahti, Finland lasted from 06/14/2000 to

06/21/2000. During this time, LAI, FPAR, canopy reflectance and transmittance and soil

53

reflectance were measured. This campaign made use of one AccuPAR, two LAI-2000s,

one ASD and one LI-1800.

A 1x1km area was chosen as the sampling site. This area was divided into a grid with

each cell measuring 50x50m. Based on CCD images, the canopy was divided into three

different classes, open area (100x150m), regular forest (100x150m) and dense forest

(200x200m). For each of these classes, a single sub-plot was created, with cells being

divided into even quarters. LAI and canopy reflectance were measured at every site,

while FPAR and soil reflectances were only measured in the open area, and canopy

transmittance only in the regular forest area.

54

6 CONCLUSIONS

The general objective of this research was to contribute to the post-launch validation of

the MODIS LAI / FPAR product. The specific focus was to show the validity of the

MOD15A1 (daily) product, its conglomeration into the publicly released MOD15A2 (8-

day) data set, and thereby validate our production of a MOD15A3 (monthly) product, to

be made available from the Climate and Vegetation Group’s ftp/web site.

Statistical analysis of the product was limited to days before September 13, 2000, due

to a problem with changing definitions of the QA field. For the data sets analyzed, the

product from the main LUT algorithm behaves in the general manner governed by the

physics of the problem and other constraints discussed earlier. The back-up NDVI

algorithm, being a based on a single measure, is incapable of distinguishing realistic

inputs from bad inputs. Therefore, errors randomizing the inputs will also randomize the

outputs. For this reason, there is no agreement between the values produced by these two

algorithms for a given pixel. While an additional source of possible uncertainty is

removed by the algorithm’s independence from the MOD35 cloud mask product, the end

user should not rely on the explanation of unmarked clouds for every occurrence of main

algorithm failure. Separating these events from other causes would be a line of future

investigation.

Analysis performed on field data has thus far shown the operational product to be

within the limits of uncertainties of ground truth measurements. Due to the large

granularity of the MODIS pixel, the network of global campaigns provides information

on an insignificant fraction of operational product. Because of the long-term goals of the

55

MODIS produced data set it will be important to maintain on-going field validation to

watch for systematic errors that could possibly be induced by changing geolocation

information or other low level inputs used in the processing higher order products.

56

REFERENCES

Berthelot, B., Adam, S., Kergoat, L., Cabot, F., Dedieu, G., and Maisongrande, P. 1997.

A global dataset of surface reflectances and vegetation indices derived from

AVHRR / GVI time series for 1989 1990: The LAnd SUrface Reflectances (LASUR)

data, in Proc. Commun. 7th Int. Symp. Physical Measurements and Signatures in

Remote Sensing. Courchevel, France, Apr. 7-11, 1997.

Brown, L. J., J. M. Chen, S. G. Leblanc, and J. Cihlar. 2000. Short wave infrared

correction to the simple ratio: an image and model analysis. Remote Sensing of

Environment, 71:16-25.

Campbell, J.L., Burrows, S., Gower, S.T. and Cohen, W.B. 1999. BigFoot Field Manual

Version 2.

Carlson, T. N., and Ripley, D. A. 1997. On the relation between NDVI, fractional

vegetation cover, and leaf area index. Remote Sensing of the Environment, 62:241-

252.

Chen, J. M. and J. Cihlar. 1995. Plant canopy gap size analysis theory for improving

optical measurements of leaf area index. Applied Optics, 34:6211-6222.

Chen, J. M. 1996. Canopy architecture and remote sensing of the fraction of

photosynthetically active radiation in boreal conifer stands. IEEE Transactions on

Geoscience and Remote Sensing, 34:1353-1368.

57

Deschamps, P.Y., Bréon, F.M., Leroy, M., Podaire, A., Bricaud, A., Buriez, J.C., and

Sèze, G. 1994. The POLDER mission: Instrument characteristics and scientific

objectives. IEEE Transactions on Geoscience and Remote Sensing, 32:598-615.

Diner, D.J., Asner, G.P., Davies, R., Knyazikhin, Y., Muller, J.P., Nolin, A. W., Pinty,

B., Schaaf, C.B., and Stroeve, J. 1999. New directions in Earth observing: Scientific

application of multi-angle remote sensing. Bulletin of the American Meteorological

Society, 80:2209-2228.

Gobron, N., Pinty, B. and Verstraete, M. M. 1997. Theoretical limits to the estimation of

leaf area index on the basis of visible and near-infrared remote sensing data. IEEE

Transactions on Geoscience and Remote Sensing, 35:1438-1445.

Hansen, M.C., DeFries, R.S., Townshend, J.R.G., and Sohlberg, R. 2000. Global land

cover classification at 1km spatial resolution using a classification tree approach.

International Journal of Remote Sensing, 21:1331-1364.

Jasinski, M. F. 1996. Estimation of subpixel vegetation density of natural regions using

satellite multispectral imagery. IEEE Transactions on Geoscience and Remote

Sensing, 34:804-813.

Justice, C. O., Wolfe, W., El-Saleous, N., Descloitres, J., Vermote, E., Roy, D., Owens,

J., Masuoka, E. 2000. The Availability and Status of MODIS Land Products. The

Earth Observer, 12:8-18.

Justice, C. O., Vermote, E., Townshed, J.R.G., Defries, R., Roy, D.P., Hall, D.K.,

Salomonson, V.V., Privette, J.P., Riggs, G., Strahler, A., Lucht, W., Myneni, R.B.,

Knyazikhin, Y., Running, S.W., Nemani, R.R., Wan, Z., Huete, A., Leeuwen, W.,

58

Wolfe, R.E., Giglio, L., Muller, J.P., Lewis, P. and Barnsley, M.J. 1998. The

moderate resolution imaging spectroradiometer (MODIS): Land remote sensing for

global research. IEEE Transactions on Geoscience and Remote Sensing, 36:1228-

1249.

Knyazikhin, Y., Glassy, J., Privette, J.L., Tian, Y., Lotsch, A., Zhang, Y., Wang, Y.,

Morisette, J.T., Votava, P., Myneni, R.B., Nemani, R.R., and Running, S.W. 1999.

MODIS Leaf Area Index (LAI) and Fraction of Photosynthetically Active Radiation

Absorbed by Vegetation (FPAR) Product (MOD15) Algorithm, Theoretical Basis

Document, NASA Goddard Space Flight Center, Greenbelt, MD 20771, USA.

Knyazikhin, Y., Martonchik, J.V., Diner, D.J., Myneni, R.B., Verstraete, M., Pinty, B.,

and Gobron, N. 1998b. Estimation of vegetation leaf area index and fraction of

absorbed photosynthetically active radiation from atmosphere-corrected MISR data,

Journal of Geophysical Research, 103:32239-32256.

Knyazikhin, Y., Martonchik, J.V., Myneni, R.B., Diner, D.J., and Running, S.W. 1998a.

Synergistic algorithm for estimating vegetation canopy leaf area index and fraction

of absorbed photosynthetically active radiation from MODIS and MISR data.

Journal of Geophysical Research, 103:32257-32274.

Leroy, M., Deuze, J.L., Breon, F.M., Hautecoeur, O., Herman, M., Buriez, J.C., Tanre,

D., Bouffies, S, Chazette, P., and Roujean, J.L. 1997.Retrieval of atmospheric

properties and surface bidirectional re ectances over land from POLDER/ADEOS.

Journal of Geophysical Research, 102:17023-17037.

59

Liu, Q., and Huete, A. 1995. A feedback based modification of the NDVI to minimize

canopy background and atmospheric noise. IEEE Transactions on Geoscience and

Remote Sensing, 33:457-465.

Lotsch, A., Tian, Y., Friedl, M.A., and Myneni, R.B. 2001. Land cover mapping in

support of LAI and FAPAR retrievals from EOS-MODIS and MISR: Classification

methods and sensitivities to errors. International Journal of Remote Sensing, (in

press).

Loveland, T.R., Zhu, Z., Ohlen, D.O., Brown, J.F., Redd, B.C., and Yang, L. 2000. An

analysis of the IGBP global land cover characterization process. Photogrammetric

Engineering and Remote Sensing, 65:1021-1031.

Morisette, J., Privette, J., Guenther, K., Belward, A., and Justice, C. O. 2000. The CEOS

Land Product Validation (LPV) Subgroup: Summary of May 23-25th Meeting.

Earth Observer, 12(4):6-9.

Myneni, R.B., Hall, F.G., Sellers, P.J., and Marshak, A.L. 1995. The meaning of spectral

vegetation indices. IEEE Transactions on Geoscience and Remote Sensing, 33:481-

486.

Myneni, R.B., and Ross, J. 1991. Photon-Vegetation Interactions: Applications in Optical

Remote Sensing and Plant Physiology, Springer-Verlag, New York, USA.

Myneni, R.B., Hoffman, S., Knyazikhn, Y., Privette, J.L., Glassy, J., Tian, Y., Wang, Y.,

Song, X., Zhang, Y., Smith, G.R., Lotsch, A., Friedl, M., Morisette, J.T., Votava,

P., Nemani, R.R., and Running, S.W., submitted April 2001. Global Products of

60

Vegetation Leaf Area and Fraction Absorbed PAR from Year One of MODIS Data.

Remote Sensing of the Environment.

Panferov, O., Knyazikhin, Y., Myneni, R.B., Szarzynski, J., Engwald, S., Schnitzler,

K.G., and Gravenhorst, G. 2001. The role of canopy structure in the spectral

variation of transmission and absorption of solar radiation in vegetation canopies.

IEEE Transactions Geoscience and Remote Sensing, vol. 39, No. 2, February 2001.

Price, J. C. 1993. Estimating leaf area index from satellite data. Remote Sensing of the

Environment, 31:727-734.

Privette, J.L., Myneni, R.B., Morisette, J.T., and Justice, C.O. 1998. Global validation of

EOS LAI and FPAR products. The Earth Observer, 10:39-42.

Ross, J. 1981. The Radiation Regime and Architecture of Plant Stands. Dr. W. Junk

Publ., The Hague, The Netherlands.

Ross, J., Knyazikhin, Y., Marshak, A., and Nilson, T., Mathematical Modeling of the

Solar Radiation Transfer in Plant Canopies (in Russian, with English abstract). 195

pp., Gidrometeoizdat, St. Petersburg, Russia, 1992.

Sellers, P.J., Randall D.A., Betts A.K., Hall F.G., Berry J.A., Collatz G.J., Denning, A.S.,

Mooney H.A., Nobre C.A., Sato N., Field C.B., Henderson-sellers A. 1997.

Modeling the exchanges of energy, water, and carbon between continents and the

atmosphere. Science, 275:502-509.

Shabanov, N.V., Zhou, L., Knyazikhin, Y., Myneni, R.B., Tucker, C.J., submitted Aug.

2000. Analysis of inter-annual changes in northern vegetation activity observed in

61

AVHRR data during 1981 to 1994, IEEE Transactions on Geoscence and Remote

Sensing.

Tian, Y., Zhang, Y., Knyazikhin, Y., Myneni, R.B., Glassy, J., Dedieu, G., and Running,

S.W. 2000. Prototyping of MODIS LAI and FPAR algorithm with LASUR and

Landsat data. IEEE Transactions on Geoscience and Remote Sensing, 38:2387-

2401.

Tian, Y., Wang, Y., Zhang, Y., Tabor, K., Hoffman, S., Privette, J., Knyazikhin, Y.,

Myneni, R.B. 2000. Documentation of Kalahari Transect Measurements in

Botswana, March 2000.

Tian, Y., Wang, Y., Zhang, Y., Knyazikhin, Y., Bogaert, J., and Myneni, R.B. submitted

April 2001. Radiative Transfer Based Scaling of LAI/FPAR Retrievals from

Reflectance Data of Different Resolutions. Remote Sensing of the Environment.

Wang, Y., Tian, Y., Zhang, Y., El-Saleous, N., Knyazikhin, Y., Vermote, E., and

Myneni, R.B. 2001. Investigation of product accuracy as a function of input and

model uncertainties: Case study with SeaWiFS and MODIS LAI/FPAR algorithm.

Remote Sensing of the Environment, (in press).

White, H. P., Pavlic, G., Hoffman, S., Myneni, R., Latifovic, R., and Chen, J.M. 2000.

Land Cover and LAI Mapping of Kejimkujik for August 2000. Proceedings,

Kejimkujik Fall 2000 Conference, Halifax, Nova Scotia.

Woodcock, C.E. and Strahler, A.H. 1987. The factor of scale in remote sensing. Remote

Sensing of the Environment, 21:311-332.

62

Zhang, Y., Tian, Y., Knyazikhin, Y., Martonchik, J., Diner, D., Leroy, M., Myneni, R.B.

2000. Prototyping of MODIS LAI and FPAR algorithm with POLDER data over

Africa. IEEE Transactions on Geoscience and Remote Sensing, 38:2402-2418.