**Land Cover Change Detection in Southern Brazil Through Orbital Imagery Classification Methods**

José Maria Filippini Alba, Victor Faria Schroder and Mauro Ricardo R. Nóbrega *Brazilian Agricultural Reasearch Corporation (Embrapa) Brazil* 

### **1. Introduction**

98 Remote Sensing – Applications

Vila, D.A.; de Goncalves, L.G.G.; Toll, D.L. & Rozante, J.R. (2009). Statistical Evaluation of

Continental South America. *J. Hydrometeor*. Vol.10, 533–543.

Combined Daily Gauge Observations and Rainfall Satellite Estimates over

Remote sensing has been considered a promising technology as support for agriculture since its beginnings, due to its contribution for a climatic perspective or for understanding of processes related to land. However, significant applications occurred only in the late twentieth century, as result of the creation of best orbital systems, with higher spatial resolution, more bands and stereoscopic capture. Several orbital platforms, as AQUA/TERRA, Quickbird and Ikonos are examples in that sense (Moreira, 2005; Embrapa, 2009).

Engineering innovations, new sensors and methods of digital image processing must be performed simultaneously so that the advances in remote sensing will be achieved. Anyway, the incorporation of orbital images on geographic information systems (GIS) and their post-processing appear as significant application since a daily life perspective, specially when classification methods are involved, because of their relation to land use, land cover and easy interpretation.

This chapter considers classification methods applied on orbital imagery in Southern Brazil, in the coastal plain of Rio Grande do Sul state (Fig. 1), where a sequence of lagoons and lakes of different sizes occurs in the context of subtropical to temperate climate with cold winters and hot summers, being organized according to the following four sections:


All the exposed data are related to research projects of the Embrapa Temperate Climate Research Center, Pelotas, Rio Grande do Sul state, one of the 45 research centers of the Brazilian Agricultural Research Corporation (Embrapa) spread on the national territory.

#### **2. About classification methods**

Classification methods were created in the statistical context, when a collection of objects or samples could be characterized and separated in different classes (Davis, 1986). The method

Land Cover Change Detection in Southern

Brazil Through Orbital Imagery Classification Methods 101

Spectrometric methods measure the response of target materials in the laboratory or field. Then the spectral patterns are simulated for a specific sensor through a specialized software, so that a sequence of orbital images is classified according to the pattern generated by that

Classification of remote sensing images appear as useful tool in terms of land use, whether in local scale or in regional scale. Filippini-Alba and Siqueira (1999) classified land use in the municipality of Pelotas, Rio Grande do Sul state, Brazil, according to nine classes: agriculture, clay soils, forestry, natural forest, pastures, soil without vegetal cover, urban, water and wetlands. Natural forest and pastures occupied 23% and 30% of the territory respectively, with intense interference between the classes "urban" and "soil without vegetal cover". Similar classes were considered by Bolfe et al. (2009) for land use in Rio Grande do Sul state, but with different results. Agriculture and pastures occurred 32% and 50% of the territory respectively with only 3% for natural forest. The differences between both studies are easy explained in term of scale, because in the two occasions Landsat images were used and a municipality was considered at the former and a state at the latter,

Lu et al. (2004) discriminated seven categories of change detection techniques: (i) algebra; (ii) transformation; (iii) classification; (iv) advanced models; (v) Strategies with geographic information systems (GIS); (vi) visual analysis and (vii) miscellanea. The classification methods are detached, with six different modalities. One of them, the "Post-classification comparison", is predominantly used in this chapter. That is, multi-temporal images are classified separately into thematic maps, then the classified images are compared pixel by pixel. The "Post-classification comparison" minimizes the atmospheric impacts, the environmental differences among multi-temporal imagery, as well as differences related to the sensor kind, providing a complete matrix of change information. However, some disadvantages can be appointed, because a great amount of time and expertise is required and, by other side, the final accuracy depends on the quality of the classified image due to

Guild et al. (2004) quantified the areas of deforestation in the Amazonian forest, state of Rondonia, Brazil. The tasselled cup transformation (Crist & Kauth, 1986) was applied with the Landsat imagery from the years 1984, 1986 and 1992. The variables brightness, greenness and wetness were evaluated for each year, then, a file integrated the nine levels of information (three variables by three years). These data were processed through principal components and classification methods with overall accuracy of 79.3, 68.4% and 71.4%, for tasselled cap land cover change classification, tasselled cap with principal components land cover change classification and tasselled cap image differencing, respectivelly. Final classes were a combination between land cover and time, so change detection was quantified.

The two applications present in this chapter consider the Supervised classification method with maximum likelihood as criteria for definition of the classes. The proximity of the study areas and knowledge of the territory justify this option to take advantage of available information. Unsupervised classification is a fast proccess, good for unknown or outlying areas, when truth of field is unavailable and most time-consuming after processing, due to the need of class identification. Maximum likelihood criteira is restricted by software and time-consuming but it represents a improvement in relation simple criteria as the

software (Lillesand & Kiefer, 1994; Pontara, 1998; Moreira, 2005).

with territory difference of 1 to 155 times in size.

the weather condition on that date.

parallelepiped or the minimum distance.

Fig. 1. Location of the areas of study in the context of South America, Brazil and the State of Rio Grande do Sul. Montenegro municipality as gray levels composite of the bands 1, 4 and 5 TM/Landsat 5 (Miranda, 2002, 2005) and Caiuba lagoon in gray levels composite of the bands 3, 2 and 1 ASTER (NASA, 2004) on september 26th, 2000.

was extended for processing of digital images considering the pixels as objects to be classified (Crosta, 1993, Lillesand & Kiefer, 1994, Jensen, 1996). The application of the classification methods on satelite imagery is affected by two main factors:


The "supervised classification" is included in topic (i), when the user defines the groups through digitalization of uniform spectral answer. Statistics are calculated for each group, so the classification is performed for all the image (pixel by pixel). The unsupervised classification eliminates the intervention of the user; then, the software defines the groups by means of scattergrams "band versus band", where isolines related to distribution density of the pixels are analysed (Crosta, 1993).

Different options for case (ii) are posible, by instance, criteria can use the standard deviation for the parallelepiped method or a minimum distance when an eliptical form is defined for each group or a combination of the later with statistical probability, that is, the maximum likelihood method. Jensen (1996) presented other criteria of classification, as Isodata method and the Fuzzy method.

Fig. 1. Location of the areas of study in the context of South America, Brazil and the State of Rio Grande do Sul. Montenegro municipality as gray levels composite of the bands 1, 4 and 5 TM/Landsat 5 (Miranda, 2002, 2005) and Caiuba lagoon in gray levels composite of the

was extended for processing of digital images considering the pixels as objects to be classified (Crosta, 1993, Lillesand & Kiefer, 1994, Jensen, 1996). The application of the

The "supervised classification" is included in topic (i), when the user defines the groups through digitalization of uniform spectral answer. Statistics are calculated for each group, so the classification is performed for all the image (pixel by pixel). The unsupervised classification eliminates the intervention of the user; then, the software defines the groups by means of scattergrams "band versus band", where isolines related to distribution density

Different options for case (ii) are posible, by instance, criteria can use the standard deviation for the parallelepiped method or a minimum distance when an eliptical form is defined for each group or a combination of the later with statistical probability, that is, the maximum likelihood method. Jensen (1996) presented other criteria of classification, as Isodata method

bands 3, 2 and 1 ASTER (NASA, 2004) on september 26th, 2000.

i. Intervention of user.

and the Fuzzy method.

ii. Criteria of definition for the groups.

of the pixels are analysed (Crosta, 1993).

classification methods on satelite imagery is affected by two main factors:

Spectrometric methods measure the response of target materials in the laboratory or field. Then the spectral patterns are simulated for a specific sensor through a specialized software, so that a sequence of orbital images is classified according to the pattern generated by that software (Lillesand & Kiefer, 1994; Pontara, 1998; Moreira, 2005).

Classification of remote sensing images appear as useful tool in terms of land use, whether in local scale or in regional scale. Filippini-Alba and Siqueira (1999) classified land use in the municipality of Pelotas, Rio Grande do Sul state, Brazil, according to nine classes: agriculture, clay soils, forestry, natural forest, pastures, soil without vegetal cover, urban, water and wetlands. Natural forest and pastures occupied 23% and 30% of the territory respectively, with intense interference between the classes "urban" and "soil without vegetal cover". Similar classes were considered by Bolfe et al. (2009) for land use in Rio Grande do Sul state, but with different results. Agriculture and pastures occurred 32% and 50% of the territory respectively with only 3% for natural forest. The differences between both studies are easy explained in term of scale, because in the two occasions Landsat images were used and a municipality was considered at the former and a state at the latter, with territory difference of 1 to 155 times in size.

Lu et al. (2004) discriminated seven categories of change detection techniques: (i) algebra; (ii) transformation; (iii) classification; (iv) advanced models; (v) Strategies with geographic information systems (GIS); (vi) visual analysis and (vii) miscellanea. The classification methods are detached, with six different modalities. One of them, the "Post-classification comparison", is predominantly used in this chapter. That is, multi-temporal images are classified separately into thematic maps, then the classified images are compared pixel by pixel. The "Post-classification comparison" minimizes the atmospheric impacts, the environmental differences among multi-temporal imagery, as well as differences related to the sensor kind, providing a complete matrix of change information. However, some disadvantages can be appointed, because a great amount of time and expertise is required and, by other side, the final accuracy depends on the quality of the classified image due to the weather condition on that date.

Guild et al. (2004) quantified the areas of deforestation in the Amazonian forest, state of Rondonia, Brazil. The tasselled cup transformation (Crist & Kauth, 1986) was applied with the Landsat imagery from the years 1984, 1986 and 1992. The variables brightness, greenness and wetness were evaluated for each year, then, a file integrated the nine levels of information (three variables by three years). These data were processed through principal components and classification methods with overall accuracy of 79.3, 68.4% and 71.4%, for tasselled cap land cover change classification, tasselled cap with principal components land cover change classification and tasselled cap image differencing, respectivelly. Final classes were a combination between land cover and time, so change detection was quantified.

The two applications present in this chapter consider the Supervised classification method with maximum likelihood as criteria for definition of the classes. The proximity of the study areas and knowledge of the territory justify this option to take advantage of available information. Unsupervised classification is a fast proccess, good for unknown or outlying areas, when truth of field is unavailable and most time-consuming after processing, due to the need of class identification. Maximum likelihood criteira is restricted by software and time-consuming but it represents a improvement in relation simple criteria as the parallelepiped or the minimum distance.

Land Cover Change Detection in Southern

historical point of view.

similar way.

were adjusted by visual observation.

the area occupied by the class "Pastures".

Brazil Through Orbital Imagery Classification Methods 103

instantaneous field of vision (IFOV) of 79 meters and 240 meters respectively. Improvements of the system included more bands (short-medium infrared) and reduction of IFOV to 30 meters and 120 meters respectively, for the thematic mapper (TM) in 1982 (Jensen, 1996). A panchromatic band was developed for the Landsat 7 satellite, with the TM plus sensor, but, the series reached to the end. The Landsat 5 satellite was an engineering success, the platform was launched in 1984 and is still on orbit. Anyway, the Landsat series represents the greatest collection of terrestrial images for environmental applications, specially, since a

Composites of three bands were used, with green band (500 – 600 nm), red band (600 – 700 nm) and near infarred band (700 - 800 nm) for the MSS sensor and the red band (630 – 690 nm), the nearinfarred band (760 – 900 nm) and the shortmedium infrarred band (1550 – 1750 nm) for TM sensor. These games of bands are not equivalent, then similar patterns of colour

Digital imagery was registered for the Universal Transverse of Mercator projection(UTM), zone 22 South with the datum WGS84, after that, a mosaic of pairs of scenes was composed, by instance, scene 237/82 and 237/83 for MSS sensor. So the mosaic was cutted evolving the study area and a file with the mentioned three bands was created for each date. Initially, data were processed by the supervised classification according to the maximum likelyhood criteria. Eigth poligons of homogeneous features were digitalized with the software ER-Mapper (1995), deriving in the test areas, then each pixel of the corresponding image was classified according to its similarity with the parameters of each test area (beach/dunes, forestry, rice crops, pastures, sandy fields, soil without vegetal cover, water and wetlands). A second strategy was developed to improve results, so the "potential area for agriculture", that is rice crops, pastures and soil without vegetal cover, was isolated and classified by

Results of the preliminar process of classification considered a rectangle of 30 km wide and 65 km long for the images of 2001, 2002 and 2005 (Fig. 2). The "potential zone for agriculture" is represented by a "central zone" in direction south - north to the East of Merin Lagoon, where agricultural areas are discriminated. A confusion between rice crop class and wetlands class is observed in the west - north sector of the study area. Sandy fields are long structures related to old movements of the sea (Atlantic Ocean), where a low charge of livestock is a common use and forestry is developed eventually, as observed in the images. The area occupied by water bodies was almost constant, that is 19 - 21% (Table 2), but, the wetlands were reduced in area in 2005, a year of drought probably, then, there was an increment in the area occupied by the class "Soil without vegetal cover" and a reduction of

When the "potential zone for agriculture" was isolated, the precison of evaluation of the area occupied by pastures, rice crops and soil without vegetal cover (SWVC) was improved. The kind of sensor, the date of the image and the meteorological conditions induced diferences among the imagery of different dates (Fig. 3). The images of 1973 and 1981 present a different characteristics due to captation with the MSS sensor. The first image corresponds to september, when the culture had not been implanted yet. Some agricultural areas showed different pattern in 2001 and 2002 (same harvest) related to waterlogged soils, probably, due to intense rain in that time. A differencial answer of the vegetation in the agricultural areas was observed since 2005, what suggests a evolution of the vegetal development of the rice

Acording to Lu et al. (2004), methods (iii) and (v) were considered in this chapter. Classification (iii) was applied in both conditions, Caiuba lagoon and Montenegro municipality. The extraction of the poligon corresponding to the "potential area for agriculture" in the vicinity of Caiuba lagoon represents a tipical strategy of GIS (v). Softwares of digital images processing and GIS are very similar. Both can execute multilayer processing, including raster/vector files and logic/mathematical algorithms, but digital images processing is more specific for raster format and GIS for vector format.
