GIONAL
VEI 7RU,
1062
legislation
utability of
using both
imagery.
"ark (NNP)
applied to a
i conjugate
‘erability of
ng process,
ixels of the
ferent areas
inced ANN
Over.
'esetzlichen
die hat die
» analysiert.
0 IKONOS
nerkundung
en für acht
et. Die
pfehlungen
te mit dem
ten um 15%
rten sich zu
id Probleme
e.g. Landsat
ver, such as
^ dditionally
plicable 10
f mapping
|]ti-temporal
> imagery in
p vegetation
own, 2000).
n IKONOS
algorithms,
, has been
80%. This
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
paper will report on the analysis of the suitability and
transferability of Artificial Neural Networks (ANN), using both
mixed and unmixed pixels, as a remote sensing algorithm for
mapping and monitoring land cover at regional areas together
with the high spatial resolution of multispectral Carterra'™ Geo
[KONOS imagery.
1.2. Artificial Neural Networks in Remote Sensing
Artificial Neural Networks have been applied to several remote
sensing studies often resulting in higher or equal mapping
accuracies than achieved with traditional classification
methodologies or mixture modelling (Benediktsson et al., 1990;
Foody et al., 1995; Atkinson et al, 1997). An advantage of
ANNS is the ability to generalise, and they do not require end-
member spectras for soft classifications approaches (Lippmann,
1987; Heppner et al., 1990; Atkinson and Tatnall, 1997; Foody
et al, 1997). It has also been found that ANN require less
training data than traditional remote sensing classification
approaches, such as Maximum Likelihood classification
(Heppner et al., 1990). However, remote sensing applications
of ANNs have tended to use simple data sets consisting of few
pixels and land cover classes (Bernard, 1998). The
generalisation ability of ANNs has been found to be limited, as
ANNs tend to be overfitted to the training data (Wilkinson,
1997). Additionally, the fine scale spectral variation and low
number of pixels available for training in comparison to the
number of pixels within the image have caused difficulties in
applying ANN to high spatial resolution imagery. In previous
studies, the transferability over large geographical areas has
been found to be limited and dependent on the ability of the
ANN to generalise (Egmont-Petersen et al., 2002). This study
therefore aims to examine the ability of ANNs to transfer
trained knowledge, acquired by a classifier on one area, to
classify unseen data across large geographical areas and
potentially multi-temporal imagery.
2. METHODOLOGY
2.1. Study area
The study area was located in the Northumberland National
Park (NNP) in Northern England (UK). The NNP is one of the
eleven English and Welsh National Parks and is valued for its
biodiversity and wildlife. It is covered mainly by upland
vegetation, such as bracken, heather moorland (approximately
20% of England’s upland vegetation resources) and blanket bog
(approx. 18 % of England’s resource) (ERDP, 2000). These
resources are a significant proportion of the worldwide
resources, as almost 15% of the world’s blanket bog can be
found in Britain (RSPB, 2000; Backshall et al., 2001).
However, the increasing pressure of changes in the environment
and in management practices have impacted the status,
composition and extent of important vegetation habitats and
resulted in significant changes in the extent of upland
communities and their biodiversity (Tallis, 1985). The
requirement for new monitoring and management schemes and
the high spatial variation of upland vegetation made the NNP an
ideal test site for this study. The specific site covered by the
IKONOS imagery is the British Ministry of Defence’s (MoD)
Otterburn Training Area (OTA). The Otterburn range is located
in the centre of NNP and covers 229 square km, approximately
20% of the Park area.
911
2.2. Image and ground data acquisition
A Carterra'™ Geo IKONOS image, recorded on 2™ September
2002, was acquired for the majority of the OTA. The image
was georeferenced to the British National Grid (BNG) using 18
GPS ground control points, resulting in a Root Mean Squared
(RMS) error of 2.57 m. Corrections for relief displacement,
caused by the altitude range of the imagery of 500 m, were also
carried out, (Hanley and Fraser, 2001). Further details of the
geometric and atmospheric corrections applied to the image
may be found in (Mehner et al., 2003; Mehner et al., in press).
Ground survey data offered the only source of providing
information on the surface land cover distribution. Several GPS
field campaigns using the kinematic Leica GPS 500 system
were carried out along different transects across the imagery.
Sample points were recorded at least every metre yielding 3D
coordinates and the vegetation type attribute. The GPS
coordinates were transformed to BNG and thereby referenced
with the imagery. The number of measurements per class for
each pixel was used as a guide to calculate the land cover
distribution of each pixel. The number of mixed pixels applied
were however much smaller than pure pixels, reflecting the true
spatial variation of land cover found at the test site.
Two transects were located within the same geographical area
(training site), while the third transect was carried out at a site 5
km away (remote site). Both areas are similar in terms of
vegetation types and altitude and are relatively flat, thus the
spectral variation due to anisotropic reflectance effects was
minimised. Initially the ANNs were trained using data, both
mixed and unmixed pixels, from the training site and then
applied to the remote site to test the performance of ANNs
when classifying unseen data of a different geographical
location.
2.3. Artificial Neural Network design and training
considerations
The ability of an ANN to classify unseen data successfully and
thereby transfer its trained knowledge dependent on its ability to
generalise (Haykin, 1999). The generalisation of an ANN is
influenced by several parameters, which require an optimal
choice. :
2.3.1. Design The classification performance of an ANN is
influenced by its design depending on a choice of several
parameters, such as the number of hidden nodes and learning
algorithms (Haykin, 1999). The most common type of ANN
used in remote sensing is the Multilayer Perceptron (MLP),
which was also chosen for this study (Lippmann, 1987; Lees,
1996; Atkinson and Tatnall, 1997). MLPs have shown to be a
suitable ANN design for many remote sensing applications
(Lees, 1996). MLPs consist of three different kinds of layers:
input layer, hidden layer and output layer. The number of nodes
in the input layer is determined by the number of input bands,
which is: four IKONOS bands - blue, green, red and near-
infrared - as well as the Normalized Difference Vegetation
Index (NDVI), calculated to enhance the spectral separability
(Figure 1). The number of output nodes is dependent upon the
number of land cover classes in the classification scheme: in
this case eight different upland vegetation classes, such as
Calluna vulgaris, Mire and Molinia Cearula (Figure 1) were
considered.
