TRANSFERABILITY OF ARTIFICIAL NEURAL NETWORKS FOR MAPPING LAND COVER OF REGIONAL 
AREAS WITH HIGH SPATIAL RESOLUTION IMAGERY 
H. Mehner™, D. Fairbairn®, E. Csaplovics" and M. Cutler? 
* School of Civil Engineering and Geosciences (Geomatics), University of Newcastle, Newcastle upon Tyne, NEI 7RU, 
United Kingdom 
Email: henny.mehner@ncl.ac.uk, dave.fairbairn@ncl.ac.uk 
b Institut fuer Photogrammetrie and Fernerkundung, Technische Universitaet Dresden, Mommsenstr. 13, 01062 
Dresden, Germany 
Email: csaplovi@res.urz.tu-dresden.de 
* Department of Geography, University of Dundee, Dundee, DD1 4HN, United Kingdom 
Email: m.e.j.cutler@dundee.ac.uk 
Theme Session 8: Advanced Classifiers and Data Fusion Techniques 
KEY WORDS: artificial intelligence, neural networks, IKONOS, multitemporal, land cover, mapping, ecology 
ABSTRACT: 
Accurate and frequently updated land cover maps of environmentally protected areas are necessary for the management of legislation 
programs governed by the EU, national authorities and local environmental schemes. This study has analysed the suitability of 
Artificial Neural Networks (ANN) for mapping and monitoring land cover over regional areas, such as National Parks, using both 
hard and soft classification approaches together with the high spatial resolution of multispectral Carterra™ Geo IKONOS imagery. 
The study aimed to examine the transferability of remote sensing mapping algorithms over Northumberland National Park (NNP) 
located in Northern England. The ANNs were trained using ground data of eight different upland vegetation classes and applied to a 
multispectral IKONOS image of NNP. The ANNs applied consisted of a Multiple Layer Perceptron (MLP), using a conjugate 
gradient descent, and one hidden layer with a varying number of hidden nodes and combinations of weights. The transferability of 
ANNs was found to depend on the ability to generalise, which could be improved by applying early stopping in the training process, 
improving the accuracy of the validation data by an average of 15%. The classification accuracies for validation pixels of the 
training areas resulted in 80%, but decreased to less than 50% if evaluated against validation pixels acquired from different areas 
within NNP. Limitations and issues regarding the transferability of MLP ANNs were observed to be significant. Advanced ANN 
algorithms such as Support Vector Machines were required to enable the use of ANNs for mapping and monitoring land cover. 
KURZFASSUNG: 
Genaue und regelmäßig aktualisierte Karten von umweltgeschützten Gebieten sind notwendig für die Verwaltung von gesetzlichen 
Schutzmaßnahmen durchgeführt von der EU, bundesweiten und regionalen Umweltschutzprogrammen. Diese Studie hat die 
Eignung und Nutzbarkeit von künstlichen Netzwerken als Kartierungs- und Überwachungsmethode für regionale Gebiete analysiert. 
Dabei wurden softe und harte Klassifizierungsmethoden, zusammen mit hochauflösenden multispektralen Carterra ™ Geo IKONOS 
Bildern, verwendet. Das Ziel der Studie war die Verbesserung der Ubertragbarkeit von Kartierungsalgorithmen der Fernerkundung 
fiir das Gebiet des Northumberland Nationalparks im Norden Englands (GB). Die kiinstlichen Netzwerke wurden fiir acht 
verschiedene —Hochlandsvegetationsklassen trainiert und auf multispektrale IKONOS Bilder angewendet. Die 
Gewichtskombinationen und die Anzahl der Neuronen des mittleren Layers wurde nach verschiedenen Literaturempfehlungen 
variiert. Die Übertragbarkeit von künstlichen Netzwerke wurde beeinflusst von ihrer Generalisierung. Diese konnte mit dem 
vorzeitigen Beenden des Trainingvorgangs verbessert werden und die Genauigkeit der Klassifizierung von Vergleichsdaten um 15% 
erhöht werden. Die Klassifikationsergebnissen für Pixels der Trainingsgebiete erreichten um die 80%, aber verschlechterten sich zu 
unter 50% für Pixels des IKONOS Bildes von andern Gebieten des Nationalparks. Die Studie zeigte Begrenzungen und Probleme 
bei der Übertragung von künstlichen Netzwerken auf. 
until recently been limited in its spatial resolution (e.g. Landsat 
TM) in relation to the spatial variability of land cover, such as 
1. INTRODUCTION 
that of upland vegetation (Taylor et al., 1991). Additionally 
Environmentally protected areas are monitored by different remote sensing has yet to be successfully applicable to 
legislation and management programmes introduced by the EU, monitoring schemes, allowing the transfer of mapping 
national and local environmental management bodies, such as algorithms across geographical areas and multi-temporal 
national park authorities. Such management schemes require by imagery. 
legislation the accurate and frequent mapping and monitoring of The development of high spatial resolution satellite imagery i 
land cover, for example for upland vegetation found in national the last five years has offered a new potential to map vegetation 
parks in the UK. Traditional mapping approaches, such as field ^ regularly and at a more suitable scale (Slater and Brown, 2000). 
surveys and the interpretation of aerial photography have been However the classification of high spatial resolution IKONOS 
shown to be low in accuracy, time consuming and therefore imagery using traditional remote sensing mapping algorithms, 
expensive (Cherrill et al., 1994). Remote Sensing has been seen such as the Maximum Likelihood Classification, has been 
as a potential mapping methodology for the last 20 years but has limited to accuracies ranging between 52% and 80%. This 
910 
Inter: 
—— 
pape 
trans 
mixe 
mapr 
with 
[KOI 
1.2. / 
Artif 
sensi 
accul 
meth 
Food 
ANN 
mem 
1987. 
et al. 
traini 
appre 
(Hep 
of Al 
pixel: 
gener 
ANN 
1997 
numb 
numb 
apply 
studic 
been 
ANN 
there! 
traine 
classi 
poten 
2.1.5 
The : 
Park | 
elevei 
biodi 
veget 
20% « 
(appre 
resoui 
resoui 
found 
Howe 
and 
comp 
result 
comm 
requir 
the hi 
ideal 
IKON 
Ottert 
in the 
20% €