FULLY AUTOMATED, HIGH-RESOLUTION CLASSIFICATION OF REMOTELY SENSED DIGITAL 
MULTISPECTRAL RECORDINGS 
by B.-S. Schulz 
Institut für Angewandte Geodäsie, Frankfurt a.M., Germany, Com. III 
ABSTRACT 
The working program pre-set by the chairmen of WG III/3 has been worked on further rigorously as 
regards topic 1 (Analysis of Multispectral Digital Recordings). The results obtained have been 
condensed in a method of automatic classification. Thus, e.g. multispectral homogeneity and normal 
distribution of grey values within so-called training areas ensure that theoretical preconditions of 
the Maximum Likelihood Method (MLM), which was used for their discrimination, are met. From this a 
method for searching training areas automatically has been developed. Reliability of classification 
results depends decisively on that statistically equal training areas are grouped before classifi- 
cation, and that statistically not equal but neither significantly different training areas are 
grouped only after classification. The statistical test parameters needed for this purpose are 
presented and their effects described, whereby confusion matrices and classification results serve as 
examples. 
Keywords: Classification, Feature Extraction, Landsat 
1. INTRODUCTION 
Expectations placed by the different users in 
remote sensing focus mainly on aspects such as 
accuracy of content and geometry as well as on 
the fine structured variety of land use classi- 
fications. 
The confusion matrice still serves as a source 
of information on the reliability of assignment 
of re-classified training areas as well as - 
inadmissibly - a criterion of discrimination in 
the case of larger, classified areas. The 
achievable accuracies documented in this way 
give no rise to much confidence as to the 
reliability of results. A hit rate of 60-70 % 
must be considered as a poor result unworthy of 
discussion and should be sufficient reason for 
thinking about the causes. Item 1 of the working 
programme of WG III/3 served that purpose. 
1. Analysis of multispectral digital 
recordings 
1.1 Analysis of data with regard to systematic 
and random errors and their effect on 
classification 
1.2 Possibilities of data preprocessing and 
compression without loss of informtion 
1.3 Statistical requirements on training areas 
and their statistical analysis 
1.4 Statistical analysis of clusters with 
regard to separability of objects, 
admissibility of their integration before 
classification, and necessity for their 
integration after classification. 
1.5 Analysis of mainly used or self-developed 
classifiers with regard to their 
separation capability 
1.6 Analysis and valuation of classification 
algorithms and procedures as well as 
of spectral resolution of different remote 
sensing systems 
1.7 Possibilities and limits of unsupervised 
classification 
1. FIRST RESULTS OBTAINED BY A NEW PROCEDURE 
The causes of poor discriminations in the 
classification as well as first results of new 
methodological developments have been published 
(Schulz, 1990). Considering that the acquisition 
of objects and land uses can mainly be performed 
indirectly via spectra, the spectral homogeneity 
and normal distribution constitute important 
factors, that means that, among others, neither 
a variance to be pre-set may be exceeded nor may 
there exist more than one frequency maximum. 
This comparatively simple requirement is in most 
cases very difficult to be met if one pre-de- 
fines training areas externally, particularly in 
instances where they are obviously inhomogeneous 
in land use and hence also spectrally inhomo- 
geneous, as e.g. in cases of sparsely built-up 
areas, mixed forest, etc. 
2. AUTOMATIC SEARCH AND GROUPING OF TRAINING 
AREAS 
The problem of training area definition 
inadmissible in the sense as outlined above may 
be solved by scanning the data without training 
area-related a priori definitions, that means 
automatically without operator or intepreter, 
for those image sections which fulfill the a.m. 
distribution criteria in all n spectral bands 
within the range of a pre-set gliding working 
matrice. In these instances of a successful 
search for training areas their substitutional 
parameters, i.e. mean value vector and covarance 
matrice, are stored. 
The training areas found in this way do not at 
all contain à priori significantly different 
spectral qualities of land uses. For this reason 
it is important to check in the following every 
training area by comparing it to each other one 
for whether 
- it is statistically equal to another one and 
can thus be combined with it into a 
training area already before classification 
- it is significantly different from the othef 
one and must hence enter into 
classification as representing a completely 
new type of land use or 
- whether it is, in the sense as defined above, 
neither equal with any other area nor 
significantly different from it and must 
therefore first be formally introduced into