International Archives of Photogrammetry and Remote Sensing, Vol. 32, Part 7-4-3 W6, Valladolid, Spain, 3-4 June, 1999 
175 
COMBINING SPECTRAL AND TEXTURAL FEATURES FOR MULTISPECTRAL 
IMAGE CLASSIFICATION WITH ARTIFICIAL NEURAL NETWORKS 
H. He , C. Collet 
Department of Geography, University of Fribourg, Perolles, 1700 Fribourg, Switzerland 
hongchang.he@unifr.ch , claude.collet@unifr.ch 
KEYWORDS: Artificial Neural Networks, Multispectral Image, Landcover and Landuse, Classification, Grey Level Co-occurrence 
Matrix. 
ABSTRACT 
A new procedure for multispectral image classification was investigated in this study. This procedure incorporates textural features 
into conventional per-pixel classification and is implemented with artificial neural networks. The grey level co-occurrence matrix for 
textural measure calculation is derived from a supervised spectral classification based on three SPOT spectral bands (XS1, XS2 and 
XS3). Four textural measures, contrast, entropy, angular second moment and inverse difference moment, are calculated based on the 
grey level co-occurrence matrix. An artificial neural network using a variant of the back-propagation algorithm is applied to the 
procedure. 
The performance of the new procedure is assessed relative to that of a classification combining spectral features and textural 
measures based on the grey level co-occurrence matrix from the near-infrared SPOT band and a per-pixel classification procedure. 
The test results show that the classification accuracy with the new procedure can be improved by more than 10%, compared to the 
other two classification procedures. 
The study indicates that the textural measures based on the grey level co-occurrence matrix from the supervised spectral classification 
can reveal more effectively spatial forms of landcover and landuse in multispectral images. However, an inappropriate window size 
for textural feature extraction affects the fidelity of textural features. Window sizes of 5x5 and 7x7 can be considered as the 
appropriate ones for landcover and landuse classification in this test area. 
1. INTRODUCTION 
With the advent of higher spatial resolution satellite images and 
their wide application to landcover and landuse mapping, 
conventional per-pixel classifiers present more and more 
limitations for the data processing. This is mainly due to two 
factors. Firstly, conventional classifiers assume that landcover 
classes in satellite images follow some kind of specific 
distributions, e.g. Gaussian distribution, but in fact they are 
often spectrally complex and heterogeneous. Secondly, per- 
pixel classifiers are based only on spectral features, ignoring the 
spatial information in images. 
Many efforts have already been devoted to overcome these 
problems, and three procedures are representative among them. 
The first incorporates textural features into image classification 
(Haralick et al., 1973; Jensen, 1979; Marceau et al., 1990; 
Baraldi and Parmigianni, 1990; Sadler et al., 1991). However, 
this method is under dispute. Some studies indicated that it only 
provides a relatively small and sometimes even no practical 
improvement of the accuracy of landcover and landuse 
classification in urban areas compared to the use of only 
spectral features (Jensen, 1979; Baraldi and Parmigianni, 1990; 
Sadler et al., 1991). There are two possible reasons for these 
results: (a) the procedure used for extraction of textural features 
is based only on a single spectral image; thus, textural measures 
cannot completely reveal spatial forms of landcover and landuse 
in multispectral images; (b) the method also depends on 
statistical assumptions about class distributions. The second 
procedure is a spatial (or contextual) reclassification technique 
(Wharton, 1982; Gong and Howarth, 1992; Fung and Chan, 
1994). As two subsequent classifications are implemented in 
this scheme, errors from the two stages and their interaction 
would probably result in a decrease in the overall classification 
accuracy. The third procedure applies artificial neural networks 
(ANNs) in multispectral classification (Bischof et al., 1992; 
Paola and Schowengerdt, 1997). A continuously increasing 
number of investigations is focused on this classification 
procedure for three reasons: 
- The procedure does not require any a priori knowledge of the 
statistical distribution of the data, i.e., nonparametric 
classification. 
- It can classify landcover and landuse classes with any 
distribution, i.e. not only Gaussian but also multimodal or 
even disjoint. This means that the procedure is capable of 
classifying classes with heterogeneous data and high spatial 
variability. 
- It is adaptable and robust. 
Although this procedure results in high classification accuracy, 
typically more than 80%, it is also necessary to combine textural 
features into the procedure to fully exploit the potential of 
ANNs. However, very few attempts have been made up to now 
in this direction.