Dissertação
Avaliação da técnica de modelo linear de mistura espectral como subsídio à classificação do uso e ocupação do solo
Fecha
2023-03-24Autor
Fernanda Paula Bicalho Pio
Institución
Resumen
The constant development of technologies for the acquisition and processing of remote sensing
data has encouraged the application of digital images in various circumstances, including
environmental studies and monitoring of the Earth's surface. Considering the development of
anthropic actions and the intensification of alternative land use for various activities, studies of
land use based on remote sensing data have become more common and geotechnologies are
considered important tools for providing data that help territorial management work. The
application of images provided with free access also demonstrates greater use, including data
from the sentinel satellite. The multispectral images obtained through remote sensing allow the
application of analysis techniques to obtain qualitative and quantitative information on a given
region. A significant application of these data is observed in analyzes of the use and coverage
of the earth's surface, which is knowledge of great importance for territorial management. For
land use and occupation classification studies, several image processing and classification
techniques can be applied. These, in turn, vary depending on the characteristics of the scene
acquisition sensors and the purpose of the experiment. In image classification studies, it is
common to encounter spectral mixing problems that are observed due to the spatial resolution
of the sensors. Depending on the purpose and parameters of the study to be carried out, the
spectral mixture may be characterized as a limitation of the data and processes carried out.
Thus, working methods are used to extract information from images with greater detail,
considering the properties of the materials present within a pixel. The study of the spectral
mixture is used to help the classification techniques of digital images. In the analysis of the
spectral mixture, the spectrum presented inside a pixel is decomposed and different methods
can be applied for this purpose. The most frequently used method is the linear model of spectral
mixing (MLME). The ease of operation of the method makes it used in studies in different parts
of the world with good results. The present work aims to address the linear model of spectral
mixing based on the method of least squares with restriction. This model assumes that the
spectral response at each pixel in any spectral band is due to a linear combination of the spectral
responses of each component present in the mixture. Thus, in this study, we sought to compare
two classification results of land use and occupation for the Ribeirão Jirau microbasin region,
located in the municipalities of Itabira and Santa Maria de Itabira, in the interior of Minas
Gerais, with a mapping generated from from conventional classification with scenes of digital
images from the Sentinel-2 satellite and the other result generated from fraction images
obtained in the linear model of spectral mixture. The results were compared based on
concordance coefficients extracted from the error matrix, where general precision analyzes
were considered (global accuracy and Kappa coefficient) and the individualized analysis by
class with the study of user and producer precisions, seeking to know detail the performance of
the classifiers. The overall precision values for classifications derived from Sentinel-2 and
MLME data were 65% and 62%, respectively. The Kappa coefficient presented a result of 0,53
for the classification obtained from the satellite spectral data and 0,62 for the classification
obtained from the mixture model, indicating that the quality of the two mappings can be
considered as moderate, good enough or good. The analysis of the classifications by category
revealed that the two mappings presented specificities, mainly for the classes of natural forest
and planted forest. The study allowed attesting the functionality of using MLME fraction
images as a subsidy for image classification and provided possibilities for new types of analysis
for the application of the model.