MAPPING VEGETATION CHANGES IN DUTCH HEATHLAND, 
USING CALIBRATED LANDSAT TM IMAGERY. 
E.J. Van Kootwijk* and H. van dcr Voet** 
* Research Institute for Nature Management, P.O. box 9201, 6800 I IB Arnhem, The Netherlands 
** Agricultural Mathematics Group, P.O. Box 100, 6700 AC Wagcningen, The Netherlands 
Presented at ISPRS Commission VII Symposium, September 17-21, 1990, Victioria, B.C. Canada 
Abstract. Using multivariate regression techniques to establish a relation between pixel values and 
percentages ground cover, it was possible to estimate the relative cover per pixel of heather and 
grass species. By using an optimization procedure for localising field data in tne image, the quality 
of training sets was improved. 
Key Words: Regression model, image calibration, training set optimization, heathland. 
INTRODUCTION 
Air pollution causes a higher atmospheric 
input of nutrients in Dutch hcathlands. As a 
result, the rate of change from heather to 
grass species increases (Berdowski 1987, Hcil 
19841. Mapping the changing amounts of 
heatner and grass in relation to nutrient input 
serves both nature management and environ 
mental decision making. This study was carried 
out as part of the Dutch National Research 
Program on Acidification. The objective was to 
develop within one year a mapping method 
that would yield quantitative information on a 
national scale. Moreover, it had to be suitable 
for further development into a monitoring 
system. Landsat TM imagery was found to be 
the optimal choice as to availability, spatial 
resolution, area covered and monitoring poten 
tial. 
Processing satellite imagery of natural veget 
ation using classifying algorithms poses serious 
difficulties. In fact, interpreting such images is 
not a classification problem. Natural vegetation 
often exhibits a large spatial and temporal 
variability in species abundances, structure and 
lifeform. This variability is not to be seen as 
noise, but as an important property of the 
vegetation. Even heathland, thougn rather 
simple in composition, can not be described 
rightly in terms of 30 x 30 m squares (the 
resolution element of Landsat TM) containing 
either heather or grass. The alternation of 
grass (mostly Molinia cacrulca and Dcschamp- 
sia flexuosa) and heather species (Calluna 
vulgaris and Erica tetralix), sometimes in pat 
ches, sometimes truly intermingled, is an ecolo 
gical factor, related to the physiological and 
ecological processes occurring in the vegeta 
tion. Most TM pixels contain both heather and 
grass, as spatial variation is high relative to 
the size of the resolution element. Instead of 
classifying pixels, one would like to know the 
amount of various cover types that contributed 
to individual pixel values. In the absence of an 
appropriate canopy reflectance model and an 
atmospheric model, regression techniques can 
be used to model the relation between cover 
and reflectance. 
REGRESSION MODELS FOR IMAGE 
CALIBRATION 
A multivariate calibration model is needed in 
order to predict ground cover composition 
from reflection data. Denote for pixel i the 
ground cover data by the vector x, = 
(x, 1 ,...,x IK )’, where K is the number of cover 
classes, and the pixel values by the vector y, 
= (yi 1 ,...,y iq )’, where q is the number of spec 
tral bands. There are two special characteris 
tics in this situation. Firstly, the ground cover 
data x, are vectors of non-negative values 
summing to 1. Secondly, physical theory tells 
us to expect that y,j is a linear combination of 
the cover fractions x, k (at least under idealised 
circumstances). That is: 
E(yij) = «Li (1) 
where |x kj is the expected pixel value of a pixel 
consisting of 100% class k cover in spectral 
band j. 
To some extent these characteristics lead to 
conflicting requirements for the calibration 
model, because a linear model will inevitably 
be able to predict cover fractions smaller than 
0 or larger than 1, whereas any nonlinear 
model violates the expected relation (1). In 
this paper a nonlinear model is used. 
A further point deserving attention is the 
possibility of using prior information on 
ground cover composition in the population of 
ixcls to be predicted. Such information may 
e available from external sources or from the 
training data. The latter possibility is relevant 
especially when the training pixels have been 
obtained by a suitable sampling scheme (e.g. 
simple or stratified random sampling, or some 
form of cluster sampling). However, also in 
other situations the assumption is often accep 
ted that the training sample is representative 
of the whole population. In this paper the 
utility of incorporating such prior information 
is evaluated. 
The problem of predicting cover composi 
tion has been considered by several other 
authors, c.g. Marsh c.a. (1980), Switzer (1980), 
Pech c.a. (1986), Wood and Foody (1989). In 
685