lage 
ugh 
| the 
the 
: the 
ts of 
the 
tags 
d to 
, an 
etric 
arge 
ixel 
sub- 
data 
dent 
hich 
| not 
/are. 
due 
ture 
nder 
| the 
radic 
that 
the 
sing 
For 
light 
cily, 
is to 
full 
gital 
S T. 
'eam 
1g of 
t the 
Ware 
iphic 
;ucas 
and 
9.6: 
ing. 
mera 
eport 
95, 
eras: 
dings 
ation 
s of 
Mare 
  
THE FUSION OF GIS INFORMATION AND REMOTELY SENSED DATA FOR 
MAPPING EUROPEAN SCALE LAND COVER 
Christian G. Hoffmann 
European Commission 
Joint Research Centre (JRC) 
Space Applications Institute (SAI) 
Environmental Mapping and Modelling Unit (EMAP) 
1-21020 Ispra (VA), Italy 
Commission IV, Working Group 1 
KEY WORDS: GIS, Classification, Accuracy. 
ABSTRACT: 
The paper illustrates the complementary usage of low and high spatial resolution remote sensing data and GIS information in the 
production of a digital European land cover map at an approximate scale of 1:2million. Regionalization, pixel class labelling, and 
classification accuracy assessment procedures are performed at a variety of spatial scales using ancillary remotely sensed 
preclassified data and spatial and aspatial information held in a GIS. The study indicates that high spatial resolution remote 
sensing images can be used to assign and validate AVHRR clustering results. The processing procedures described in the study 
could not be implemented in an efficient, reliable or timely manner without the use of GIS techniques. 
INTRODUCTION 
The Environmental Mapping and Modelling Unit (EMAP) 
of the Space Applications Institute (SAI) has initialised a 
project to demonstrate the feasibility of producing a 
European land cover map using Local Area Coverage 
(LAC) NOAA-AVHRR data. Land cover mapping 
procedures usually require extensive user interaction, 
particularly in the assignment of meaningful land cover 
labels and in the assessment of classification accuracies. 
The cost of user interaction rapidly becomes prohibitive as 
the amount of remotely sensed data increases (Hoffmann 
and Belward, 1996). The study indicates that summary and 
agglomerated statistics of high spatial resolution remote 
sensing images can be used to assign and validate AVHRR 
clustering results. 
There is a rich history of using NOAA-AVHRR data for 
land cover classification at regional and continental scales. 
This is because of the moderate spatial resolution of the 
AVHRR sensor and its daily world-wide coverage. Most 
studies have used vegetation indices extracted from 
AVHRR time series to discriminate between land cover 
types. Recently it has been empirically demonstrated over 
Europe (Roy, 1996) and Africa (Lambin and Ehrlich, 1995) 
that the inclusion of surface temperature can discriminate 
regional land cover classes more effectively than 
vegetation indices alone. In this study surface temperature 
(Ts) information and a vegetation indices (NDVI) are 
derived from multitemporal AVHRR data. This is for two 
reasons, firstly, in an attempt to increase classification 
accuracy and secondly because there is evidence to suggest 
that NDVI and Ts may be interpreted in a biophysically 
meaningful manner at regional or continental scales. 
355 
METHODOLOGY 
NDVI and Ts data, extracted from a time series of 
European AVHRR images are used as the primary data 
source. These data are processed independently on a 
regional basis in an attempt to reduce regional variations in 
NDVI and Ts and in an attempt to automate the process as 
shown in Figure 1. Principle component analyses is 
performed in each region to reduce inherent data 
redundancies (e.g. temporal correlation) and to remove 
residual noise.) The methodology used for land cover class 
separation is unsupervised classification, followed by 
cluster labelling. The cluster labelling is performed by 
examination of NDVI and Ts values and by the 
examination of pre-classified high spatial resolution (20m) 
SPOT/LANDSAT TM test images selected from the data 
archives of the Monitoring Agriculture by Remote Sensing 
(MARS) project. In addition, urban classes are located 
across the entire image using the Digital Chart of the 
World database. 
The AVHRR clusters in each different region, resulting 
from unsupervised classification, are labelled into land 
cover types by resampling, using the nearest neighbour 
resampling scheme to the same pixel resolution as the pre- 
classified high spatial resolution test imagery. This 
permits a one to one pixel relationship between the two 
scales of data. The MARS classes are then agglomerated 
and compared with the unassigned resampled AVHRR 
clusters. This is achieved by histogramming the 
agglomerated MARS classes that fall under each unique 
AVHRR cluster. Several test images within the same 
region are treated independently in this manner and then 
compared to ensure consistency within region cluster 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B4. Vienna 1996