ing 
ined ‘ 
per. 
he 
illetin. ^ 
. Proc. 
Oliveira, Hermeson 
SEGMENTATION AND CLASSIFICATION OF LANDSAT-TM IMAGES TO MONITOR THE 
SOIL USE 
HERMESON NÓBREGA BARROS DE OLIVEIRA! 
OLGA REGINA PEREIRA BELLO? 
KLAUS DE GEUS? 
Departamento de Informática - UFPR 
Centro Politécnico , Jardim das américas 
CEP: 81531-970, Curitiba - PR 
‘hermeson @zipmail.com.br 
*olga@inf.ufpr.br 
*klaus @inf.ufpr.br 
KEY WORDS: Classification, Segmentation 
ABSTRACT: A hybrid unsupervised/supervised classification method is described and applied to 
Landsat TM images. This method differs from the conventional classification in the sense that the 
clustering algorithm is applied to a set of regions, obtained from the segmented image. The statistical 
parameters of these regions are used to classify each. After the previous step, some regions are selects to 
be training areas on a supervised classification step. A comparison is made between the pixel per pixel 
classification and the region classification. 
1 INTRODUCTION 
Know and represent the Earth surface have always been a preoccupation of many civilizations. The increase 
of complexity of localization, mapping and monitor the planet surface, with the objective of territorial limits 
protection, knowledge, use and preservation of natural resources, are responsible for the advance in the ways 
of receiving and manipulating Earth surface information. 
Brazil, with it's huge territory, characterized by a great agroecological diversity, combined with a very 
dynamic soil occupation, presence for decades an extremely disoriented populational flux. The necessity to 
know, map and monitor natural resources and populational movements, with the objective of a more 
controlled occupation of the territory and a rationally use of natural resources, have encouraged the execution 
of projects for researching and map the Brazilian surface (Bognola et al, 1997) 
Traditionally, classification techniques have been divided into two big categories: supervised and non- 
supervised, according to the proceedings used to obtain the training areas. In Brazil, classification techniques, 
are not used with the same intensity as other countries. That's because of different reasons, for example: 
processing capacity and storage of great volume of data, the necessity to train teams and the overall 
performance of these techniques. 
Some experiments were done to test segmentation methods in the Amazon area (Almeida, 1994). Although 
the results were considered satisfactory, these experiments only tested the performance in homogenous areas, 
where big properties and few variety of soil use are predominant, don't reflecting the huge amount of 
situation of soil use, that exist in Brazil. The necessity and urgency to know and monitor soil use and natural 
resources, in the whole country, turn the segmentation and classification tools to be of great interest. 
This paper has the objective of showing an hybrid methodology of Landsat-TM images classification. First, a 
region-growing algorithm is used to divide the image into spectral homogenous regions. Then, a clustering 
algorithm is applied over the segmented image. The next step is a selection of regions that will be used as 
training areas on the supervised classification step. The last step is, using the training areas, apply a 
supervised algorithm and classify the image. 
  
  
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 
1065