176
was the acquisition of Landsat TM images and work in the field with observations
and class records on the present land use and occupation; the third stage involved
the integration of data collected in the field with the data from the Landsat TM sensor in GIS environment, using the ArcGIS™ 10.1 and the ENVI™ software, followed by the final draft of the research. In the next lines we detail the second and
third stage procedure.
The Land Use and Occupation Maps (1984 and 2014) were made with a
1:100,000 scale, using object-targeted classification of orbital images. Basically,
this type of segmentation and classification consider various characteristics such as
spatial (medium) and spectral (color) heterogeneity and the difference of the surrounding objects; in other words, it uses the region’s growth method to add neighboring pixels.
Moreira (2003) states that the classification uses algorithms and the spectral patterns in the image are recognized in a sample from the training area provided by the
analyst, thus validating the need to know the area of study and improving the quality
of the map generated.
For this research, four images were used from the Thematic Mapper sensor TM,
orbit 215 point 072 and 073 and orbit 216, point 072 and 073, with satellite passing
dates of April 1980 and 2014. The TM sensor aboard LANDSAT 5 and LANDSAT
8 satellites makes images of the Earth’s surface, producing images of 185 km
(width) of the ground, with spatial resolution of 30 m and 7 spectral bands. The
satellites revisit time to capture the same portion of land is 16 days.
Orbital images were chosen considering the smallest possible amount of clouds,
lower excess of brightness and higher spectral normality. During this search, it
became clear that the region had a predominance of orbital images covered by clouds,
which somewhat limited the choice of images, especially regarding available dates.
After selection, these images were georeferenced through Envi
®
5.0 software
with the IBGE’s topographic sheet as basis, which covers the area of study, plus a
1:100,000 scale in digital format. Then images were grouped into a false-color composition (R5G4B6). The representative interpretation keys of each class of interest
were chosen for the algorithm used in the object-targeted classification. Nine classes
were defined based on the Technical Manual on Land Use of the Brazilian Institute
of Geography and Statistics (Ibge 2013): Urbanized Areas, Inland Water, Ocean
Water, Temporary Cultures, Pastures, Eucalyptus Forestry, Forest Area, Mangrove
Areas and Uncovered Areas. The use of symbols and standard forms also followed
the same manual.
The next step was generating the maps using the Envi feature extraction module
that extracts information and classifies images based on spectral, spatial and texture
characteristics. Initially, image segmentation was applied on homogeneous regions.
For the item “edge” threshold 40 was applied, and 50 for the “merge setting”, both
chosen by trial and error, until a satisfactory result was obtained through visual
observation. Among the types of object-targeted classification algorithms, there are
Support Vector Machines (SVM) and K-Nearest Neighbor (K-NN). In this study,
based on Garofalo et al. (2015), SVM were used, since it provided the best accuracy
classifications for the area of study and scale used.
S.O. Souza et al.
was the acquisition of Landsat TM images and work in the field with observations
and class records on the present land use and occupation; the third stage involved
the integration of data collected in the field with the data from the Landsat TM sensor in GIS environment, using the ArcGIS™ 10.1 and the ENVI™ software, followed by the final draft of the research. In the next lines we detail the second and
third stage procedure.
The Land Use and Occupation Maps (1984 and 2014) were made with a
1:100,000 scale, using object-targeted classification of orbital images. Basically,
this type of segmentation and classification consider various characteristics such as
spatial (medium) and spectral (color) heterogeneity and the difference of the surrounding objects; in other words, it uses the region’s growth method to add neighboring pixels.
Moreira (2003) states that the classification uses algorithms and the spectral patterns in the image are recognized in a sample from the training area provided by the
analyst, thus validating the need to know the area of study and improving the quality
of the map generated.
For this research, four images were used from the Thematic Mapper sensor TM,
orbit 215 point 072 and 073 and orbit 216, point 072 and 073, with satellite passing
dates of April 1980 and 2014. The TM sensor aboard LANDSAT 5 and LANDSAT
8 satellites makes images of the Earth’s surface, producing images of 185 km
(width) of the ground, with spatial resolution of 30 m and 7 spectral bands. The
satellites revisit time to capture the same portion of land is 16 days.
Orbital images were chosen considering the smallest possible amount of clouds,
lower excess of brightness and higher spectral normality. During this search, it
became clear that the region had a predominance of orbital images covered by clouds,
which somewhat limited the choice of images, especially regarding available dates.
After selection, these images were georeferenced through Envi
®
5.0 software
with the IBGE’s topographic sheet as basis, which covers the area of study, plus a
1:100,000 scale in digital format. Then images were grouped into a false-color composition (R5G4B6). The representative interpretation keys of each class of interest
were chosen for the algorithm used in the object-targeted classification. Nine classes
were defined based on the Technical Manual on Land Use of the Brazilian Institute
of Geography and Statistics (Ibge 2013): Urbanized Areas, Inland Water, Ocean
Water, Temporary Cultures, Pastures, Eucalyptus Forestry, Forest Area, Mangrove
Areas and Uncovered Areas. The use of symbols and standard forms also followed
the same manual.
The next step was generating the maps using the Envi feature extraction module
that extracts information and classifies images based on spectral, spatial and texture
characteristics. Initially, image segmentation was applied on homogeneous regions.
For the item “edge” threshold 40 was applied, and 50 for the “merge setting”, both
chosen by trial and error, until a satisfactory result was obtained through visual
observation. Among the types of object-targeted classification algorithms, there are
Support Vector Machines (SVM) and K-Nearest Neighbor (K-NN). In this study,
based on Garofalo et al. (2015), SVM were used, since it provided the best accuracy
classifications for the area of study and scale used.
S.O. Souza et al.