You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

Proceedings, XXth congress

ul 2004
' than
2 any
ta to
1 raw
of the
el for
se the
ng the
d have
| Yoon,
Xt were
1 of the
ind the
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B-YF. Istanbul 2004

The empirical models to link the spectral reflectance to the
field measured LAI were developed separately for each forest
type as well as for all forest types. In this study, we tried to use
several vegetation indices (SVI) as independent variables to the
multiple regression model. Three other spectral indices of
brightness (BR), greenness (GN), and wetness (WT) were also
created by the tasseled cap (TC) transformation.
To build the optimal statistical regression model to estimate
LAI for entire study area, we compared several sets of
independent variables that are subset of a few spectral
vegetation indices. Two multiple regression models were built
for each of two forest types. In addition to the forest cover map,
we also need a land cover map to validate the empirical LAI
estimation algorithm. A land cover map, in which the class
categorization is comparable to the MODIS Land cover type 3-
scheme (9 classes), was obtained by ordinary maximum
likelihood classification method. Although the major portion of
the study area is forest, it also includes small and segmented
agricultural areas. The LAI values for the grass and croplands
(mostly rice paddy) were adapted from the previous study by
Hong et al. (1998).
Quality assessments of MODIS LAI product
As shown in Figure 3, the operational algorithm for producing
MODIS LAI uses two MODIS land products of the surface
reflectance (MODO09) and land cover (MODI4). The 1km
resolution MODIS LAI products are produced every 8 days,
which corresponds to the maximum value composition interval
to remove cloud cover. LAI values are calculated by
mathematical inversion of a rather sophisticated canopy
reflectance (CR) model that uses MODO09 and MOD 14. If the
CR model-based main algorithm fails, a backup algorithm
based on the empirical relationship with vegetation index is
triggered to estimate LAL
The MODIS LAI image that corresponded to the date of the
reference LAI map was obtained. Since the MODIS LAI value
is separately calculated by cover type, the MODIS land cover
products of were also acquired. MODIS land products can be
directly obtained from the Earth Observing System Data
Gateway (EOS, 2003). The MODIS LAI and land cover
products supplied by the EOSDG is originally referenced by
the sinusoidal map projection. To compare with the reference
LAI map of the study site, the MODIS products were geo-
referenced to the Transverse Mercator map projection by using
the MODIS reprojection tool (MRT) software provided by
NASA. To compare the reference LAI surface with MODIS
LAI product, the reference LAI map having 28.5m pixel size
was rescaled to Ikm pixel size.
For the quality assessment of MODIS LAI data, we applied
three phases of 1) quality of input datasets, 2) the MODIS LAI
estimation algorithms, and 3) LAI value by land cover type
(Figure 3). The accuracy of the MODIS land cover product was
assessed by the reference land cover map. We also analyzed the
effects of cloud cover at each pixel location for the MODIS
reflectance data. After the validation of input datasets, we
compared the reference LAI map with the MODIS LAI data by
the estimation algorithms. Within the scene, some pixels had
LAI value from the main CR-based algorithm and the other
pixels had LAI value from the NDVI-based backup algorithm.
As LAI value is also very sensitive to vegetation types, we tried
to analyze the MODIS LAI value by different vegetation type
(forest vegetation and grass /cropland).
assessment of Estimation of LAI
For Land
Cover types Y
MODIS LAI product
Quality |
Assessment | :
of Input data | MSS MODIS
- Estimation | Reflectance Land cover
classified accuracy | (MOD09) (MOD14)
of MODIS Land |
cover ]
- Assessmentof |
Cloud effect in |
reflectance |
| Y
7 | MODIS LAI Main
Quality | Algorithm
assessment of | - 3D Canopy Radiative
MODIS LAI | Transfer Modeling
for Algorithm |
type i: Back-up
| Fal, Algorithm
| - - Empirical
| Success modeling used
Quality | "
Figure 3. Flowchart of producing MODIS LAI and
schemes of quality assessment of MODIS LAI
Two separate multiple regression models to predict LAI for the
reference map were developed for each of coniferous and
deciduous forest. To avoid over-fitting problem of too many
independent variables (five SVIs), we imposed the rule that
only three variables can be selected for each model. Table 1
shows the selected independent variables, R^ value, and root
mean squared error of each of two models developed. Although
the use of two separate estimation models requires additional
effort of classifying the forest into two species groups, it should
be a better approach to obtain more reliable LAI map. We
generated the high-resolution reference LAI map by applying
the best regression models to the radiometrically corrected
reflectance data. Using the reference land cover type map, the
coniferous and deciduous forests were extracted prior to
applying the models and the grass and croplands were given the
same LAI value that was measured in September.
Table 1. Regression models to estimate LAI over the study area

Selected independent 2 ;
Type variables (SVI) R RMSE
Coniferous RSR, BR, GN 0.8018 | 0.6289
Deciduous NDVI, RSR, WT 0.3413 | 0.4504