62
yield modelling has a number of sources of errors such as
(a) acquisition date, (b) spatial resolution of sensor (vis-a-
vis field size and crop proportion), (c) the spectral indices
used, (d) atmosphere introduced uncertainties, and (e)
sensor-to-sensor calibration differences and their long term
spectral response.
The coarse resolution NOAA AVHRR data has been
investigated for yield modelling as its repetivity provides
matching data sets over large regions for such an
application. The major analysis steps in use of this data are
geometric correction, and multitemporal composing
(weekly/fortnightly) and NDVI images are generated.
Their further use for yield modelling could take through
any of the following parameters : Peak NDVI, NOAA-
AVHRR data has been used different approaches, such as
(i) regression at critical date with all crops (Bullock, 1992),
(ii) Cumulative NDVI with linear decomposition for crop
separation (Quarmby et al.-, 1993), (iii) Multi-year
comparison with normalisation through computation of
Vegetation Condition Ind$x (VCI)
VCI = 100 x (NDVI - NDVImin) /
(NDVImax - NDVImin)
where NDVImin and NDVImax are the minimum and
maximum NDVI for that week from all the years of data.
Physical Models of Crop Growth
Crop growth models simulate growth and develop
ment of agricultural crops based on an understanding of
underlying physical and physiological variables (Bouman,
1995). The most simplest crop growth model is Monte ith’s
(1981) light interception model, which is given by,
Wd = e.S.f.dt
where, Wd = dry weight of the crop (g m-2)
e = light use efficiency factor (g J-1)
S = incoming global irradiation (Jm-2d-l)
f = fraction light interception
t = time (d)
Fraction light interception can be estimated using the
relationship,
f= {1 -exp(-K.LAI)}
where, K is the extinction coefficient for visible radiation
(0.5 - 0.8, depending on crop type). Hence the problem of
estimating crop production reduces to quantifying LAI, i
and K values, i and K values have to be derived from field
experiments, whereas remote sensing techniques can be
used to estimate LAI.
RS observations can be integrated with crop growth
models in two ways: direct data input and ‘steering’
simulation results (Bouman, 1995). In direct data input
time-series of LAI values estimated from some Vis (e.g.
WDVI) are used to replace subroutines that simulate the
development of LAI from environmental variables. This
use of measured LAI as ‘forcing function’ generally
increases the accuracy of crop growth simulation models.
In another approach RS information can be used to ‘steer’
the crop growth models by adopting the model parameters
so that the simulated values match with RS estimated
values of LAI. This is done when the number of RS
observations is limited and/or initial conditions are not
known. This approach is used to optimise simulation
models. Clevers and van Leeuwen (1995) have linked
RS information with SUCROS (Simplified and
Universal Crop Growth Simulator) crop growth model
for yield prediction of sugar beet.
RS in Post-harvest Yield Estimation
The conventibnal procedures for crop yield
estimation involve Crop Cutting Experiments (CCE)
conducted at harvest in the plots selected using a pre
designed sampling scheme using available ground data.
In these procedures, crop vigour is not used for
stratification. Use of RS data could improve confidence
level in estimated yield (Singh et al., 1992; Murthy et
al., 1996).
For remote sensing based yield assessment a
combination of radiation interception models and
canopy temperature models seems to be most desirable
(Horie et al., 1992).
RS-BASED EMPIRICAL-STATISTICAL MODELS
INDIAN EXPERIENCE
Over the past decade, a number of crop yield
forecasting models using RS inputs have been developed
and used in making crop forecasts in CAPE project.
These have been reviewed by Dadhwal and Ray (1999).
These include (a) Single date RS-based models for
wheat (Bihar, Haryana, MP, Punjab, UP), kharif rice
(Bihar, Orissa, WB), mustard (Assam, Gujarat),
sorghum (Karnataka, Maharashtra) and cotton (Gujarat,
Maharashtra). The yield variability explained by single
date RS data based models range from 6.0 per cent for
sorghum in Karnataka to 83.0 per cent for wheat in MP,
(b) The multidate approach using crop profile fits an
analytical growth curve to spectral data and derives the
following parameters: (i) time of peak, (ii) value of peak
(iii) area under the curve, (iv) full width at half
maximum of the profile, (v) average slope of the post
heading phase. These parameters are then related to
yield. Spectral profile related growth parameters derived