International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
6) Degradation of the LISS-III LAI maps into low
resolution to compare and evaluate MODIS LAI
product.
2.7 Remote sensing data analysis and image processing
LAI may be estimated at a variety of spatial scales and with
different space-borne sensors (Chen and Cihlar, 1996) using
techniques ranging from regression models to canopy
reflectance model inversions with varying successes, which
include (1) statistical models that relate LAI to band radiance
(Badhwar et al., 1986) or develop LAI-vegetation index relation
(Chen and Cihlar, 1996 and Myneni et aL, 1997), (2)
biophysical models like Price (1993), and (3) inversion of
canopy reflectance using numerical model or LUT based model
(Gao and Lesht, 1997, Qiu et al., 1998, and Knyazikhin et al.,
1998). The empirical approach is common in large area remote
sensing LAI estimation and it is used in the present study.
The complete procedure is shown in figure 1. For each site, an
area of approximately 30 km X 30 km was extracted from IRS-
ID LISS-III. Digital numbers (DN) were converted to top-of-
atmosphere (TOA) reflectance using the sensor gain, offset,
sun-angle and exo-atmospheric band pass coefficients (Pandya
et al., 2002). Surface reflectance was calculated from TOA
reflectance using the 6S-code (Vermote et. al., 1997) using
measured aerosol optical thickness (AOT) and water vapor (wv)
at site as input. Out of six IRS-ID LISS-III acquisitions (table
1), two were cloudy and not used in analysis. The reflectance
images were registered to corresponding geo-rectified images
using nearest neighbor resampling with less than 0.5 pixel root
mean square error. The fields with LAI measurements were
identified and demarcated carefully on the corresponding LISS-
Ill and Panchromatic (PAN) merged (5.8 m spatial resolution)
data. The mean NDVI was computed for each field. Site-
specific non-linear NDVI-LAI relations were developed for
each site and acquisition. The exponential and polynomial fits
were found to have higher R? (0.58-0.73) than the linear fits
(0.3-0.52). The exponential form of models was used to
generate the fine resolution LAI maps from LISS-III data for
each site and date (e.g. figure 2). These LAI images were
aggregated (figures 3(a) and 4(a)) to 1 km spatial resolution
using averaging for comparison with MODIS LAI product.
The MODIS LAI product of | km spatial resolution and
composited over an 8-day period corresponding to 10? X 10?
tiles in HDF EOS format were acquired for study area from
EROS Data Center (table 1) and these were reprojected from
original Integerized Sinusoidal projection to UTM projection.
The LAI images (figures 3(b) and 4(b)) were generated by
applying scaling factor after masking out water and urban
pixels. Quality flags supplied with MODIS LAI products were
studied and inter-comparison between LISS-1I1 LAI and
MODIS LAI was restricted to pixels pertaining to the class of
overall best quality (cloud free pixels and LAI retrieval through
RT model).
3. RESULTS AND DISCUSSION
3.1 Variability in LAI and atmospheric measurements
across sites
Information on sites, date of LAI measurements, crops covered,
range of LAI, AOT and water vapor measured are summarized
in the table 2. The LAI for the crops considered, span a wide
range corresponding to crop emergence to peak vegetative
stage. The range of LAIs across all sites/season was from 0.14
to 5.6. The range of AOT measurements collected at the time of
satellite pass across the sites was 0.16 to 0.32 at 500 nm and the
range of water vapor was 0.71 to 1.28 cm.
Table 2: Range of LAI and atmospheric parameters across
sites/dates
Site Date of Crop** LAI AOT Water
No* Ground Range at 500 vapor
Observation nm (cm)
] 02 Dec 01 w,g 0.17-3.30 0.16 0.71
27 Dec 01 w,g 0.69-4.63 0.26 1.15
21 Jan 02° W.g 0.64-3.26 - =
2 24 Dec 01 W.g.p 0.14-3.80 0.2 0.51
18 Jan 02 ^ W,g,p 1.05-5.6 = =
12 Feb 02 w,g p 1.25-4.48 0.32 1.28
* Site No: I: Indore, 2: Bhopal, ** w: wheat, g: gram, p: pea
# Overcast conditions
3.2 Site specific LAI-NDVI empirical models and validation
of MODIS LAI
After atmospheric correction, the contrast of red and NIR
increased and consequently the estimated NDVI was higher
than those computed with DN and radiance of red and NIR
(figure 2). The exponential form of models was used to generate
the fine resolution LAI maps from atmospherically corrected
LISS-IH data. The model coefficients for different sites/dates
are summarized in table 3.
Table 3. Regression models to relate LAI to NDVI derived from
LISS 3 don, Equation yen vb in) SENDVL LAN
Site Date a b R?
Bhopal 24 Dec.2001 0.5847 01225 073
12 Feb.2002 0.5197 0.1634 0.58
Indore 02 Dec.2001 0.5156 0.1341 0.65
—— 27 Dec 2001 0.4776 0.1425 0.60
The comparison between LISS-IIT derived LAI and MODIS
LAI was carried out by performing regression between LAI
estimated from MODIS (dependent) and LISS-III (independent)
and the results are summarized in table 4 and illustrated in
figure 5. The comparison indicated significant positive
correlation between LISS-IHI derived LAI and MODIS LAI
(r=0.78 for Bhopal and r=0.72 for Indore). A slope of | and 0
intercept indicate full match, while deviations show over/under
estimation. The analyses indicated an overestimation in MODIS
LAI compared to LISS-III LAI for both sites. The scale of
overestimation is quite high for Bhopal (slope: 1.98 and 2.49)
than Indore (slope: 0.74 and 1.16). Overestimation by LAI was
higher, for higher LAI estimates by LISS-III, especially in
Bhopal (figure 5 (a), (b)). The overall root mean square error of
MODIS LAI is higher for Bhopal (0.92 and 1.26) compared to
Indore (0.20 and 0.33), however Bhopal had higher range of
LAI (0.1 to 3.28 in LISS-III LAI 0.3 to 6.9 in MODIS LAI).
During the analysis, many pixels of MODIS LAI product were
observed to have LAI (5-6.9), which seems to be unrealistically
high and contrary to the ground observations. Similar trend was
observed for other dates and site. Since comparison is restricted
to only best RT-model derived pixels, the errors could arise due
to a number of factors such as wrong biome type, vegetation
dependent parameters and effect of soil background or
aggregation procedure. Analysis of MODIS land cover product
that goes as an input in MODIS LAI retrieval algorithm,
Internati
indicate
(broad |
grasses-(
error in |
when m
classes 1
20 perc
includin,
backup
MODIS
correlati
LAT pro
using the
Table
Eq
Site
Bhor
Indoi
* Nun
A study
LAI ima
model b
significa
the MOI
of MOD
observec
arise du
soil ba
aggregat
method,
this. Sir
crop, wi
vegetatic
product.
reflectan
necessar
pointed
and is as
other pr
sites an
product
I am ind
Modelli:
out this
manuscr
Agricult
of work.
P. Singh
the field
