IAPRS & SIS, Vol.34, Part 7, "Resource and Environmental Monitoring", Hyderabad, India,2002
The value of a represents the rate of growth of profile before
Tmax and PB represents the rate of decay after Tmax (X>0 and B>0).
G(t) attains peak value Gay at time Tax. Thay iS given by:
TQ ^u. GRE [3]
The value of Gmax is calculated by substituting t by T4, in
equation (2). The width of the profile (0), i.e., difference
between first and second point of inflexions (T, and T;) on the
profile, can be calculated by:
SsIb-T-0/87 J| neun [4]
Although, originally the model was defined for tasselled cap
greenness, it has been applied for NDVI spectral profile also
(Potdar, 1993; Dubey et al., 1991).
3. STUDY AREA AND DATA USED
The present study was conducted using 15 IRS 1C/1D/P3 WiFS
data acquired over Punjab and Haryana states for 2000-2001
wheat season. The fifteen dates spanning from October 2000 to
April 2001 were Oct-26, Nov-9, Nov-19, Dec-3, Dec-13, Dec-
27, Jan-20, Feb-4, Feb-8, Feb-28, Mar-5, Mar-9, Mar-17, Mar-
22 and Apr-22. The groundtruth information on land use & land
cover type and crop details was collected by in-season field
surveys. Wheat is the dominant rabi crop of the study area. The
district-wise wheat yield estimates for 2000-2001 season were
obtained from Bureau of Economics and Statistics (BES).
4. METHODOLOGY
The analysis methodology consisted of following steps:
e Geo-referencing of multi-date WiFS data
e Conversion of multi-date images from digital number to
radiances
e Radiometric normalization of multi-date WiFS data
* Classification for wheat
e Generation of district-wise NDVI differences for wheat
and built-up areas
* Use of non-linear regression for parameterising wheat
NDVI difference values
e Analysis of spectral profile parameters with crop
phenology and yield
4.1 Geo-referencing of multi-date WiFS data
To obtain a geo-referenced multi-temporal image data set, a
WIiFS image covering study area was geometrically registered
with Survey Of India topographic maps at 1:250 000 scale
using ground control points (GCPs) and second order
polynomial transformation (master geo-referenced WiFS
image). The root mean square (rms) errors in easting and
northing were within £180 m. All the subsequent WiFS images
were geometrically registered with master geo-referenced WiFS
image using GCPs, second order polynomial transformation and
cubic spline resampling. The over all registration accuracy of
364
better than + 0.5 pixels (90 m) was achieved in each
registration.
4.2 Conversion of DN image to radiance image
In order to obtain radiometrically comparable apparent spectral
radiance data, the digital values of both the bands (Red & NIR)
have to be corrected. Using the appropriate values of calibration
constants i.e., saturation radiance (L..) and offset (Lnin)
(Ramkrishanan R, personal communication), the multi-date
dataset was converted into radiance values and then Normalized
Difference Vegetation Index (NDVI) images were generated.
4.3 Radiometric normalisation
The effect of non-target signals due to atmospheric conditions,
illumination, viewing geometry, etc., were minimised using an
image-based approach. In the present study, built-up area NDVI
subtraction method (Taylor et al., 1993) was used to normalize
the multi-date data. Built-up areas were taken as Pseudo-
Invariant Features and mean NDVI values over built-up areas
were subtracted from NDVI image to generate NDVI difference
(ANDVI = NDVIyneat = NDVIpuitup) images for all the dates.
This method provides the first order radiometric normalization
to the multi-date dataset for additive noise.
4.4 Classification for wheat crop
A hierarchical classification procedure (Oza et al., 1996) was
adopted for multi-date classification. Firstly, the data loss and
the non-vegetative classes of land use/land cover like snow/ice,
water bodies, built-up areas, barren waste lands, cloud and hill
shadows etc. were masked out by thresholding band DN and
NDVI. The non-crop vegetated areas like forest, plantation,
grasslands, and shrubs were masked out in the second step. In
the last step, the remaining pixels in agricultural areas were
assigned to various crops based on the knowledge of temporal
spectral behaviour of crops in the study area (from field surveys
conducted during the season).
4.5 Extraction of district-wise wheat NDVI
The district boundaries were overlaid on classified image and
the mean NDVI and ANDVI values for all the pixels belonging
to wheat crop in each district were computed. The district-wise
multi-date datasets of these NDVI and ANDVI values were
used for spectral profile modelling. Due to inter-mixing of
wheat and sugarcane crops in Yamunanagar district of Haryana,
relatively poor classification accuracy could be achieved; hence,
Yamunanagar was dropped from further analysis.
4.6 Estimation of spectral profile parameters
The crop emergence ANDVI (G,) value for all study districts
was taken as 0.10. This was based on a preliminary comparison
with different G, values. In other studies (Potdar, 1993;
Kaluburme et a/., 1997) a constant Tg was assumed (taking
constant Gy is more realistic than taking constant To). The sums
of squares of errors were minimized using the Levenberg-
Marquart iterative method to fit Badhwar growth profile model
(Equations 1 & 2); where t is time (in Julian days) from 1* Jan
of the year of sowing i.e. 1* Jan 2000.
The significance of goodness of fit was tested by r^ and root
mean square error (rmse) values. In addition, the significance of
estimated profile parameters was analysed in terms of standard
