NG
rs.
les were
ea using
for the
nalyze a
e within
during a
ification
tion was
ghts the
acterized
eters.
UTILIZATION OF LINEAR MIXING MODEL APPLIED TO LANDSAT-TM
DATA TO CHARACTERIZE BRAZILIAN AMAZON FOREST
Sergio Bernardes
Instituto Nacional de Pesquisas Espaciais - INPE
Caixa Postal 515, CEP 12201-970
Sao José dos Campos - SP, Brazil
ISPRS Commission VII / Working Group 3
ABSTRACT
The necessity to provide periodical studies of the amazon region, characterizing its natural resources and
anthropic alteration processes is a source of several Remote Sensing studies, many of them applying
digital image processing techniques. Conventional methods of image classification underline,
predominantly, in the spectral characteristics of the pixels, understanding them as composed by a single
class of land cover. Usually, a digital number results from integration of the responses of many targets in
the ground. In this way, the signal produced by the combination, in one pixel, of two or more classes of
land cover will not representative of none of them, resulting in a misunderstood classification. Therefore
the spectral mixture is a limiting factor in a automatic classification approach. The aim of this work is to
evaluate the use of synthetic images, obtained by a Linear Mixing Model, to characterize Brazilian
amazon vegetation. The study area consists of approximately 690 km? of the Brazilian amazon, situated
in the forest/savana ("cerrado") contact region, between 11R00'S and 51R00'W to 52R30'W. For the
methodology implementation, a visual interpretation of Landsat-TM data was performed, identifying
classes of land cover (forest, second growth forest, savanna, bare soil, ...). A Linear Mixing Model was
applied to generate three synthetic images ("vegetation", "soil" and "shade"). These images will be
classified using a maximum likelihood algorithm. The product of this approach will be compared with
the visual interpretation in a geographic information system, generating an error matrix. Kappa
coefficient of agreement will be used to determine the classification accuracy obtained with the
application of this methodology. In this way, this work intends to contribute to future space-time analysis
of the large amazon region, estimating deforesting and monitoring land occupation.
027