385
data shows
:lassification
cative of the
ogies at this
d the oilseed
s low and the
r, is not yet
oung plants.
the spring
as and beans.
emerged in
soil class.
ve occured in
fusion is now
ereals, grass
nd beans; and
racy of 51.5%
al confusion,
d the oilseed
usion between
eet, peas and
confused with
to the lower
lanting. The
the spring
nfusion. The
ively limited
11 not found,
of the winter
ise to some
g planted a
op growth and
However both
1 established
lassification
and 73.6% for
and now has a
coincident
still showing
oodland class
he deciduous
efined with a
now that the
paration from
esult in good
sses. A mean
strates this,
ique spectral
other class,
100%, though
hese were all
lassification
ed the same
onfusion was
ry pixels was
entified with
growth and
good spectral
d beans which
anges in crop
of confusion
the winter
accuracy of
een the two
lassification
ing date have
n between the
assified with
ey having the
barley has a
two woodland
0% accuracy.
As expected the multitemporal combination of data
from April and May gives the best results. Overall
class purity is now 69.78%. Classification purity is
greatly increased for the cereal crops. The unique
spectral response of the oilseed rape in May plus the
reasonable separation of the two winter cereals in
April combined with the good separation of the spring
crops in May give class purities of 100% for the
oilseed rape, 87.4% for spring barley, 82.9% for the
winter wheat and 69.7% for the winter barley. The
rather low figure for the winter barley highlights
the difficulties of accurate separation of crops with
closely matched phenologies. Inclusion of
Multispectral data from late June and early July
would in all probability resolve the winter wheat,
winter barley confusion because of the difference in
the time of crop scenescence.
Comparison of the classification crop area estimates
with proportions for the whole county and the
parishes show large errors in the classification of
the cereal crops. In particular the spring barley is
wildly over predicted. Much of the spring barley
class seems to be from boundary pixels, and so can be
considered to be a mixed pixel class. Chhikara, 1984
has shown that classification of mixed pixels is
biased and that bias in the crop proportion estimate
is increased by an increase in the relative size of
the mixed pixel class. Such bias may account for the
over estimation of the spring barley as this would
appear to be a mixed pixel class of considerable
size.
Most of the other crop types also show considerable
error in area estimation when compared with the June
census data. This tends to show that the
classification accuracy as predicted through the
confusion analysis of the training areas is rather
over optimistic. It must however be remembered that
the crop area estimates for the county, the parishes
and the independant crop area estimates used to
weight the confusion matricies are based on the June
census returns. Work as early as 1955 by Coppock
highlights the fact that inaccuracies can emerge in
such data due to the relationship between parish and
farm boundaries, and more recently Wright, 1985, in
his work on oilseed rape classification found errors
in excess of 100% in the area recorded by the census
return and the actual ground conditions. The possible
errors from this source may go part of the way
towards explaining the rather poor crop area
estimations.
5. CONCLUSIONS
The thesis that the use of multitemporal Landsat data
in crop classification will give greater
classification accuracy than single date has been
shown to hold for the United Kingdom.
The classification accuracies as expressed through
the analysis of the training data are generally
satisfactory though problems in good discrimination
between the winter cereals and grassland remain, even
when multitemporal data is used. This is a problem
associated with lack of suitable imagery. Landgrebe,
1974, and Baur et al, 1979 both report decreases in
classification accuracy through the use of mismatched
multitemporal data. However, the data set used shows
the clear advantages of using multitemporal data,
though image aquisition corresponding to the
scenescent phase of the winter cereal crops may well
prove to be vital in the successful separation of
these closely related crops.
Alternative approaches to the analysis of
multitemporal data may also improve the
discrimination between the various cover types. Work
by Richards (1984), has shown the value of using
principal components transformation in the analysis
of multitemporal data. Regions of localised change in
constant cover types were found to be enhanced in the
higher components. Simple image addition through the
use of multitemporal Landsat colour composites has
also been shown to be of value in change detection.
Areas of change showing up as a series of colours and
areas of no change showing as black and white (Eyton,
1983). Vector change detection has also been
sucessfully applied to multitemporal Landsat data.
Engvall et al, 1977, showed good proportion estimates
for winter wheat through the analysis of the temporal
trends in the Landsat mean vectors from training
areas. The problems of mixed pixels associated with
field boundaries encountered in this study calls for
the greater use of geographic information systems.
The integration of map data such as field boundaries
could considerably reduce the detrimental impact of
mixed pixels on classification accuracy.
The potential and the problems of crop identification
and monitoring from multitemporal satellite imagery
are illustrated in this study. With new developments
in this area and improved frequency of satellite
overpass allowing the aquisition of sufficient cloud
free data many of the problems will be overcome and
the potential for crop mapping and monitoring will be
increased
REFERENCES
Anderson J.E. (1985), The use of Landsat 4 MSS
digital data in temporal data sets and the
evaluation of scene to scene registration accuracy.
Photogramm. Eng. and Remote Sens.51:457-462.
Badwhar, G.D. (1984), Automatic corn-soybean
classification using landsat MSS data. 1. Near
harvest crop proportion estimation. Remote
Sens.Environ.14:15-29.
Bauer, M.E. (1975), The role of remote sensing in
determining the distribution and yield of crops.
In Advances in Agronomy (N.C.Brady,Ed.), Academic
Press, New York, 2:271-304.
Baur M.E., Cipra J.E.,Anuta P.E. and Etheridge J.B.,
(1979), Identification and area estimation of
agricultural crops by computer classification of
LANDSAT MSS data. Remote Sens. Environ.8:77-92.
Card D.H. (1982), Using known map category marginal
frequencies to improve estimates of thematic
map accuracy. Photogramm. Eng. and Remote
Sens.48:431-439.
Carlson R.E. and Aspiazu C. (1975), Cropland acreage
estimates from temporal, multi-spectral ERTS-1
Data. Remote Sens. Environ.4:237-243.
Chhikara R.S. (1984), Effect of mixed (boundary)
pixels on crop proportion estimation. Remote
Sens. Environ.14:207-218.
Coppock J.T. (1955), Farm and parish boundaries.
Geographical Studies.2:12-26.
Engvall J.L., Tubbs J.D. and Holmes Q.A. (1977),
Pattern recognition of Landsat data based
upon temporal trend analysis. Remote Sens.
Environ.6:303-314.
Eyton J.R. (1983), Landsat multitemporal color
composites. Potogramm. Eng. and Remote
Sens.49:231-235.
Hay C.M. (1974), Agricultural inventory techniques
with orbital and high altitude imagery. Photogr.
Eng.40:1283-1293.
Landgrebe D.A. (1974), A study of the utilsation of
ERTS-1 data from the Wabash river baisin, final
report, Purdue University, Purdue, IN, LARS, NASA
contract NAS 5-21773.
Lindenlaub J.C. and Davis S.M. (1978), Applying the
Quantitative approach. In (Swain P.H and
Davis S.M. Ed's) Remote sensing:the quantitative
approach. McGraw-Hill. pp.291-336.
Morain, S.A. and Williams D.L. (1975), Wheat
production estimates using satellite images. Agron
J. 67:361-364.
