Symposium on Remote Sensing for Resources Development and Environmental Management / Enschede / August 1986 
Detection of subpixel woody features in simulated SPOT imagery 
Patricia G.Foschi 
School of Geography, Oxford, UK 
ABSTRACT: A method for detecting small woody features in digital imagery was developed. Woody features, ranging 
in size from hedgerows to strips or patches several trees wide, were tested. The classification method correct 
ly detected all subpixel woody features larger than hedgerows and identified about 20 percent of the hedges. 
1 INTRODUCTION 
In Britain, changes in farming practices since the 
1940s have resulted in tree, pond, and hedgerow remov 
als, larger field sizes, and less frequent crop rota 
tions (Body 1982, Sturrock & Cathie 1980). Although 
tree and hedge removals are small land cover changes, 
they are significant landscape and habitat changes. 
These changes which decrease habitat diversity affect 
animal and bird populations. Studies have shown that 
birds and mammals utilize diverse landscape elements 
(Pollard & Relton 1970, Wegner & Merriam 1979) and 
that their distributions are related to the shape, 
size, and spatial arrangement of these landscape ele 
ments (Forman & Godron 1981, Helliwell 1976). 
Monitoring and quantifying agricultural change is 
necessary for effective land use planning and wildlife 
habitat management. Computer-assisted methods using 
remotely sensed data could provide timely monitoring 
of changes in woody vegetation which affect scenic 
beauty and wildlife. 
Since linear woody features are often subpixel tar 
gets or features smaller than the pixel size of the 
image, pixels containing these features are usually 
mixed pixels or pixels containing two or more land 
cover classes. Because the spectral values of mixed 
pixels containing woody features frequently do not 
correspond to the spectral values of woody vegetation, 
conventional multispectral classification techniques 
which operate on single pixels are problematic and re 
peatedly result in mixed pixels being placed in con 
stituent or extraneous classes. A method for detecting 
subpixel woody features in digital imagery was devel 
oped. Unlike conventional classification techniques, 
this method incorporates information about adjacent 
classes and mixture phenomena at the individual pixel 
level. 2 
2 DATA AND STUDY SITE 
As the method developed is concerned with detecting 
small landscape features, it was appropriate to use 
digital imagery of a high spatial resolution. Since 
SPOT data was not available when the project was be 
gin, simulated SPOT data was acquired for the project. 
In 1984, the National Remote Sensing Centre (NRSC) 
in Famborough, England organized a campaign to in 
vestigate the usefulness of SPOT imagery prior to its 
availability. Simulated SPOT data was collected over 
a wide variety of sites in the United Kingdom in order 
to test a number of applications (NRSC 1985)■ This im 
agery was also sold to the public and, subsequently, 
a scene was acquired for this project. Of the 39 test 
sites imaged, the Winchester data flown on 6 July 1984 
was selected because it is representative of agricul 
tural lands in lowland Britain and because it contains 
numerous linear woody features. 
A subscene of the Winchester image was then selected 
for use in developing and testing algorithms. This 
subscene, approximately 17 sq km on the ground, is lo 
cated southeast of the city of Winchester in a gently 
rolling area of mixed farmland and woodland. 
Panchromatic photography, commissioned by the Plan 
ning Department of the Hampshire County Council, was 
used to locate and map woody vegetation within the 
subscene. This photography was flown by Meridian Air- 
maps Limited on the evening of 28 July 1984 at 1:10000 
scale. Four categories of woody vegetation were 
mapped: hedgerows, single trees, single rows of trees, 
and denser woody features. Specific species were not 
identified. The airphoto interpretation was checked by 
surveying parts of the study site on the ground. 
2.1 Comparison of real and simulated SPOT imagery 
The simulated imagery was flown by Hunting Geology and 
Geophysics Limited with a Daedalus DS-1268 multispec 
tral scanner. Daedalus channels 3 through 7 were used 
singly or in combination to simulate the SPOT channels 
(Hunting Geology and Geophysics Ltd. 1984). The wave 
lengths of these simulations do not exactly match 
those of the real SPOT bands. The effect of these dif 
ferences is not known. 
Table 1. Comparison of SPOT and simulated SPOT chan 
nels . 
Channel 
SPOT 
wavelengths 
in microns 
Daedalus 
channels 
Simulated SPOT 
wavelengths 
in microns 
SI 
O.5O - O.59 
3 
O.52 - O.60 
S2 
O.6I5 - 0.68 
4+5 
O.605 - O.69 
S3 
0.79 - 0.89 
7 
O.76 - O.9O 
p 
O.5I - 0.73 
3+4+5+6 
O.52 - O.75 
The spatial resolutions of real and simulated SPOT 
data are the same: 20m in the three multispectral 
bands and 10m in the panchromatic band at nadir view 
ing. 
Since the simulated SPOT imagery was flown by high- 
altitude aircraft, it is less map-accurate than sat 
ellite imagery. Distortions in aircraft imagery are 
caused by changes in aircraft altitude and angular 
orientation during scanning. Since spatial fidelity is 
not important in this project and since techniques for 
geometric correction inevitably involve interpolation 
which further "mixes" the information in the pixels, 
there has been no attempt to geometrically correct the 
imagery. 
Another difference between real and simulated SPOT 
data is the sun angle. The simulated imagery was col 
lected at about mid-day and, consequently, shadows are