EVALUATION OF CONVENTIONAL AND CONTEXTUAL CLASSIFICA TION
TECHNIQUES TO ASSESS LAND COVER AND LAND USE IN MULTISPECTRAL IMAGES
Laerte Guimaräes Ferreira Junior
Eduardo Delgado Assad
Heleno Bezerra
Lucimar Moreira
CPAC/EMBRAPA
Laboratório de Física Ambiental
BR 020 - Km 18 - Rod. BSB/Fortaleza
73301-970 - Planaltina, DF, Brasil
ISPRS Commission VII / Working Group 3
ABSTRACT
The possibility for producing land use and land cover maps through satellite images has been increased by the
use of automatic image classification techniques. On the other hand the introduction of contextual classifiers
as well as classifiers based on image segmentation greatly improved the accuracy of the classified images.
These classifiers, on the contrary of per pixel classifiers such as the maximum likelihood and the minimum
Euclidian distance , take into account not only the spectral information related to specific pixels, but also the
location of these pixels as well as their surrounding spectral characteristics.
In order to evaluate the performance of these new classifiers implemented in the SPRING geoprocessing
system, contextual and segmentation based classifiers as well as the conventional classifiers were applied on
both enhanced (linear contrast stretching and decorrelation) and raw LANDSAT-TM data. The selected area
of approximately 62.5 km2, located in an environmental protection area near Brasilia - DF, comprises a
typical urban-rural fringe, where the natural savanna woodland is being transformed into agricultural, cattle
raising and re-forestation areas. The results already gathered have shown that the use of enhancement
techniques increases the spectral separability of all scene's targets, leading to very similar results no matter
which classifier is being used. On the other hand, when classification is not preceded by spectral
enhancement, best results concerning pattern recognition are shown by the contextual maximum likelihood
classifier.
032