selected major growing districts. In contrast to CPF, a sample
segment-based approach of 10x10 km segments and 10 per cent
sampling fraction was adopted for crop inventory. After its
operation in VIIth Five Year Plan Period, a review by Ministry
of Agriculture recommended that area contributing to 80 per
cent of production for each crop be studied. Thus, since
beginning in 1992, an aggregated area of ca. 390 Mha is being
studied yearly for issuing crop forecasts (Table 1).
CROPS PHASE -I PHASE - II
No. Study Area |No. Study Area
States |(Mha) States |(Mha)
Wheat 5 35.26 7 106.87
Rice(K) 2 28.55 11 133.98
Sorghum(R) 1 4.77 3 33.90
Mustard 4 12.88 8 2721
Groundnut(K) 1 2.86 5 36.46
Cotton 6 17.60 11 51.24
TOTAL 101.92 389.65
Table 1. Summary of study crop and region information for
Phase-I and II of CAPE project.
2.1 Crop Inventory
The CAPE procedure is being continuously revised and
upgraded to improve upon accuracy and timeliness of crop
estimates. The crop inventory experience has been reviewed by
Dadhwal et al., (2002). These efforts are related to improving a)
Sample design, b) Ground truth data collection, c) Optimising
date of data acquisition, d) Including data from additional
spectral regions in the digital analysis, e) Multi-date data
analysis, f) Use of higher spatial resolution data, g) Adopting
different classification procedures, and h) Use of microwave
data for crop inventory in kharif season. Significant results
from these studies are summarized below.
2.1.1 Sample Design : Studies carried out on rice for
Orissa and wheat for Haryana have clearly shown that two-step
stratification based on agro-physical and crop distribution
improves efficiency of stratification (Dadhwal et al, 1991;
Panigrahy et al, 1991). Size of the sample segments and
sampling fraction are two important aspects of sampling design.
Smaller sample segment sizes of 7.5 x 7.5 km and 5 x 5 km and
use of larger sampling fraction have improved district-level
estimates. Conventionally, previous season's FCC prints were
used as stratification base and year-to-year fluctuations were
not accounted. Digital stratification studies carried out for
mustard crop reduced the sampling error (coefficient of
variation of estimate) and improved efficiency of stratification
for mustard in Rajasthan (Sankhla et al., 1997).
2.1.2 Ground truth data collection: Extension of training
signatures collected at a place to some other distant areas within
the stratum does not provide satisfactory results. Locating
ground truth sites within each segment and using them for the
classification of that segment only has lead to improvement in
accuracy.
2.1.3 Acquisition Date: In a site having wheat, gram,
mustard and lentil in Hisar (Haryana), large variation in
classification accuracy due to acquisition date has been
demonstrated (Dadhwal et al., 1989, Dadhwal et al., 1996). Use
of early acquisition for wheat acreage estimation in comparison
to peak vegetative growth stage has given underestimation in
Haryana (Dadhwal et al, 1990). The early stage data are
IAPRS & SIS, Vol.34, Part 7, "Resource and Environmental Monitoring", Hyderabad, India, 2002
314
Wheat Forecasts : HARYANA
^ e
e e
Relative Deviation (96)
e
e
-15.0 -
200 | a Production |
1985 1988 1991 1994 1997 2000
YEAR (Harvest)
associated with higher inter-field as well as intra-field spectral
variability (Ruhal et al., 1988), which adversely affect crop
discrimination using supervised MXL approach. In case of rice
acreage estimation in Orissa using NOAA-AVHRR data (1989-
90 season), acquisition of October 21 gave estimate closest to
DES with earlier acquisitions having overestimates and later
acquisitions underestimates. The latter is due to progressive
ripening and harvest of rice (Panigrahy et al. 1992). In case of
wheat and mustard acreage estimation in Rajasthan, late
December/early January acquisition are optimal for mustard
with large errors in wheat estimates while February acquisition
is optimal for wheat and mustard acreage is underestimates
(due to ripening) (Purohit et al., 1997)
2.1.4 Sensor Characteristics: The spectral, spatial and
radiometric characteristics of the sensor determine crop
discrimination to a large extent. The narrow bands of TM and
LISS-I and LISS-II can lead to higher classification accuracy
than MSS. The low radiometric resolution of MOS MESSR (6
bit) leads to a lower classification accuracy than MSS and
LISS-I (Dadhwal and Parihar, 1990).
Higher accuracy of acreage estimate with TM in comparison to
MSS has been noted for rabi sorghum in Solapur in 1987-88
season (Potdar et al, 1991). Higher inventory accuracy at
village level for cotton has been noted for LISS-II in
comaprison to LISS-I in Haryana (Dutta et al., 1994). Higher
crop classification accuracy using three band data only with
SPOT in comparison to TM has been noted by Sahai et al.
(1989) in Gujarat. Singh et al., (2002) studied IRS LISS-I (72.5
m), LISS-II (36.5 m), LISS-III (23.5 m) and WiFS (188 m) and
have reported Kappa of 0.89, 0.88, 0.91 and 0.82 in wheat-gram
area. The use of higher spatial resolution can sometimes lead to
lower classification accuracy than moderate resolution
especially when training sites and locations are kept constant
due to increased variance of training sites (Dadhwal and
Parihar, 1988; Medhavy et al, 1993) Markham and
Townshend (1981) have shown that with higher spatial
resolution, spectral heterogeneity increases leading to higher
overlap between classes and decreased classification accuracy
but proportion of boundary pixels reduces which leads to
reduced misclassification. The final accuracy obtained is
‘dependent on relative importance of these two factors but
cannot be predicted since it is also dependent in a complex way
on relative location of categories within the feature space'.
Inclusion of new spectral region having additional information
is useful in crop discrimination. Thus, classification using MIR
bands show higher crop separability in wheat, gram and
mustard growing region (Dadhwal et al., 1989). Similar results
ol
ca
Vi
