degradation
rphology and
for grazing
Symposium on Remote Sensing for Resources Development and Environmental Management / Enschede / August 1986
The use of multitemporal Landsat data for improving
crop mapping accuracy
i microlight
itrated with
? use of the
/graphs.
study also
; time. These
photo inter-
very clear
ilation with
, which were
/hole system
L lot got more
irs of flight
¡rtical B&W
of which 900
300 oblique
.cal photog-
id were taken
>uld even in
>nably well,
pith overlap
I areas, or
flight, the
id a width of
1 of areas to
ap.
ferent land
; these land
e surveys it
een techni-
s well in a
capacity of
s concluded
le platform
large wing
ird genera-
performance
utilization
odelcalcula-
the future.
ults lead to
nd utiliza
il measures
:ope of this
river regime
;o virtually
L of the
iterfering
1 on the
lat may lead
)n schemes
j
floodplain,
utilization
results of
when land
:ms of the
Alan S.Belward & John C.Taylor
Cmnfield Institute of Technology, Bedfordshire, UK
Six Landsat MSS scenes from between October 1979 and August 1980 were located for a test site in southern
England. Each scene was independantly registered to the National Grid. Resampling to 50m pixel size was
carried out with Bilinear interpolation. Ground error for each scene was less than 100m. These data were
then matched to 2000ha of ground information giving the crop type and field boundaries for the 1979 - 1980
growing season. Ten cover classes were identified; Winter Wheat, Winter Barley, Spring Barley, Oilseed
Rape, Grassland, Sugar Beet, Peas, Beans, and Deciduous and Coniferous Woodland. Spectral coincident plots
were drawn for all cover types from each image, and a decision tree applied to identify key bands / dates
for maximum spectral separation of cover types. Supervised maximum likelihood classification was then used
to produce crop classifications from both single date and multitemporal data.
Classification accuracy was variable for the different cover classes. The multitemporal data gave better
overall classification accuracies than the single date images. The best result was from a spring / early
summer combination giving a mean classification purity of 70%. This is a 6% increase over the best single
date classification from May, and 46% better than the worst from February.
1. BACKGROUND
The large area coverage and sequential nature of
Landsat Multispectral Scanner (MSS) data and the
opportunity for computer data processing offers the
potential for relativly cheap, timely and accurate
crop inventory (Baur, 1975). The Landsat MSS has been
employed sucessfully for example, in crop inventory
for land use stratification (Hay, 1974), the
identification and area estimation of winter wheat
(Morain and Williams, 1975), automatic corn-soya bean
classification (Badwhar, 1984) and as an inventory
system for agriculture in California (Wall et al,
1984). Such work has involved both single date and
multitemporal image sets.
The accuracy with which individual crop types can be
classified from Landsat data varies widely. Figures
range from 80% accuracy of test field recognition
reported by Baur et al, 1979, in their work on the
identification and area estimation of corn and soya
bean, to an overall crop classification accuracy of
less than 50% found by Taylor et al, 1983, working on
crop classification in the United Kingdom.
The work by Baur et al shows that the 80% accuracy of
test field recognition was achieved by using Landsat
data for a 3 county area of Illinois, United States
of America. In this region 81% of the total land area
was cropped, and 71% of this crop land was planted
with either corn or soybean. The aquisition date of
the imagery used was from August, (identified
elswhere in the same work as being the best date for
spectral separation of the corn and soybeans).In
contrast the 50% crop classification accuracy
obtained by Taylor et al was from imagery aquired for
Feltwell, Suffolk, United Kingdom. This is an area of
mixed cropping in which a range of cereal, oilseed
and root crops are produced in small, irregularly
shaped fields. Image aquisition was also not idealy
matched to the crop calendar. Imagery was used from
April, where considerable overlap in crop spectral
response exists for this region.
Experimental proceedure may go part way to explaining
the large difference in reported classification
performance, but the contrasting agricultural
situation and match between image aquisition and
agricultural clalendar are probably of greater
importance. Carlson and Aspiazu, (1975), endorse this
view stating that, "Satellite coverage critically
timed with a crop development calendar is noted to
improve classifier effectiveness." The potential for
further improvements through the use of the temporal
dimension in crop surveys from space platforms has
long been recognised, (Steiner, 1970), though there
are few published cases where the technique has been
applied for crop inventory. Those investigators who
have used the technique have generally found that
improvements in crop classification accuracy are
found over the use of single date images. Von Steen
and Wigton, (1976), found that overall classification
accuracy of grass, cotton, corn and soybeans using
three observation dates increased by 88% over the
50.8% classification accuracy from the best single
date image. Other work such as that by Carlson and
Aspiazu, (1975), shows similar benefits from the use
of multitemporal Landsat data for cropland acreage
estimates.
More recent work (Odenweller and Johnson, 1984) has
used Landsat derived temporal-spectral profiles for
crop identification, where a temporal-spectral
profile is defined as "The multitemporal
representation of Landsat data, optimally in the form
of a green vegetation indicator, from a single
labeling target". In their work Kauth and Thomas
greeness component values were calculated for a
series of targets and plotted against time as a
temporal spectral profile. Individual cover types
were then identified within the context of baseline
proceedure analysis logic. Using this approach corn
and soybean, could be separated from perennial crops
and bare soil, then identified individually because
of differences in the amplitude and shape of the
Temporal - spectral profiles.
Badwhar, again working on corn / soybean
classification has developed an automatic
corn-soybean classification from Landsat MSS using
multitemporal data (Badwhar, 1984). This requires at
least two image aquisitions which again are
transformed to the Kauth and Thomas greeness