COMPARISON OF SOME TEXTURE CLASSIFIERS 
Einari Kilpela and Jan Heikkila 
Helsinki University of Technology 
Institute of Photogrammetry and Remote Sensing 
02150 Espoo 15, Finland 
ABSTRACT 
A performance analysis between different textural feature descriptors in land-use classi 
fication is presented. Both satellite and aerial images are used. 
The texture descriptors used are first order statistics, second order (cooccurrence) 
statistics, Fourier spectrum, amplitude varying rate statistics and fractal descriptors. 
The technical implementation of each of these descriptors in the context of classifica 
tion is also addressed. Each data set is classified using the spectral features, each 
of the texture descriptors and some variations of them, and using a combination of spect 
ral and textural features. The classifiers used are the maximum likelihood classifier 
assuming multinormal density functions, the k-NN classifier and the average learning 
subspace (ALSM) classifier. 
The performance analysis, which is based on independent test sites, shows that the ALSM- 
classifier and the k-NN classifier work equally well, but the crude assumption of normal 
densities in the context of maximum likelihood classifier produces biased results. No 
clear distinction between the behavior of the different texture descriptors was found. 
The full usage of the cooccurrence statistics works well. However, its computational 
load is quite heavy. The more simple texture descriptors, like the simple fractal di 
mension in combination with spectral features, works often equally well in the context 
of satellite images. In case of larger scale images, the more complex texture descrip 
tors are called for. 
1. INTRODUCTION 
Texture is an important cue for understand 
ing and discriminating in natural images. 
However, it is surprisingly seldom utilized 
in the context of terrain classification. 
In many applications (e.g. land-use clas 
sification) texture features could bring 
more discriminatory information. This is 
especially true when larger scale imagery 
is used. When computing textural feature 
vectors for each pixel according to some 
local neighborhood, a little bit of the 
rude assumption of spatial independence 
can be broken down. This does not mean 
that one should give up from the attempts 
to more properly model the sampling pro 
cess, e.g. with the Markov random field 
models (see /GemGem84/) and further devel 
op their computational characteristics 
especially for multi-dimensional spaces. 
This should be the final goal. In this 
paper we are anyhow concerned with more 
conservative and practical approaches. 
The problem of texture analysis and mo 
delling is a widely discussed problem in 
the areas of Pattern Recognition, Image 
Analysis, Computer Vision and even in Com 
puter Graphics. Texture is a commonly 
used criteria in the early processing of 
visual information. Paradoxically however, 
because of its loose definition, a huge 
amount of methods, both ad hoc and formal, 
have been developed (for surveys see /Ha- 
rali79/, /GoDeOo85/ and /Harali86/). The 
methods fall into two main categories, 
namely statistical and structural. The 
naming convention is slightly misleading, 
because usually quite a lot of statistics 
is involved in the structural approaches, 
too. Images taken over natural terrain 
contain both spatially and spectrally quite 
irregularly distributed, usually microscop 
ic, texture elements. The smaller the ima 
ging scale, the less structure it has. In 
many circumstances, just a simple measure 
of the roughness of the texture can bring 
enough discriminatory power to the feature 
space. Anyhow, the larger the scale, the 
more structure is visible in the texture. 
Excluding manmade objects, the spatial 
distribution of the (maybe invisible) tex 
tural structure elements is usually quite 
irregular also in large scale (aerial) 
images. Due to these facts, statistical 
methods are preferred when analyzing textu 
res in natural images. So is the case 
also in the underlying project. 
Because of the variability of the texture 
measures, a practitioner faces the problem 
of choosing the most suitable descriptor 
for his application. Reviewing the litera 
ture does not help much, because no tho 
rough comparison exists. There are so many 
factors which influence the performance 
of a texture classifier (the data, the 
texture descriptor, the number of features, 
the type of classifier, the number of 
training samples, resolution level, prepro 
cessing steps etc. ) that a complete compa 
rison would be a huge task. There are some 
texture measures, which have been quite 
successful in single comparative studies 
and which have become quite popular. One 
of the most popular texture descriptors 
is the second order statistics (Cooccur 
rence Statistics), originally suggested 
by Haralick, Shanmugam and Dinstein in 
1973 (/HaShDi73/). Another, widely used 
descriptor is the Fourier power spectrum. 
These two methods are compared in many