**5. Method**

For better understanding, the proposed method is divided into five distinct phases. The first stage concerns data collection and image pre-processing. In the second phase there is NDVI calculation and the isolation of band 4 (near infrared - IR). The creation of routines that detect fires using two test images and that is based on a threshold T is performed in the third phase. The information on the existing biodiversity in the UUES and the preprocessing and data collection steps occur in the fourth phase. Finally, the fifth phase concerns analysis of the results based on the quantification of burned areas and climatic phenomena.

#### **5.1 Obtaining and pre-processing the data**

All the nine images were geometrically corrected by means of image registration. The reference (georeferenced) image was also provided by INPE, at no cost, and the control points were collected in this image. Landsat TM images are first corrected by INPE, but with approximate coordinates, which simplifies the record of these scenes. For this reason, a firstdegree polynomial model and the interpolator "nearest neighbor", which maintains the original Digital Values DV (or Gray Levels), were used. Since the digital values define the different features in digital images, depending on the intensity of the changes of the digital values, it could be difficult to carry out fire analyses.

#### **5.2 Preparing the NDVI and IR images**

The spectral signature of the "burned" feature varies according to the fire status at the moment of image collection. If there is a fire, the digital values in the green (G) and red (R) bands are medium or close; and in the infrared band (IR), the DV are low. Though, when there is no fire, the DV in the IR band is null, spectral characteristics of the ash and coal. In the G band, the DV are low, being however larger than the values found in the R band. In this context, the pixels of the "burned" class can be classified by vegetation indexes, which are based on the difference between the DV of the IR and R bands.

NDVI (Normalized Difference Vegetation Index) – index is selected because of its efficiency in the identification of burned areas. The method is widely recognized in the literature and uses simple calculations. This index is calculated by equation 1

$$\text{NDVI} = \text{(IR-R)} / \text{(IR+R)}\tag{1}$$

The isolation of the IR band (band 4 - TM) is a simple task, which is performed by ensuring the geometrical correction and cut out over the area of interest. There are no alterations in the original digital values in this band. Fig. 4 illustrates the image IR band and the resulting image of the NDVI calculation, both originated from the image shown in Fig. 3.

The simultaneous use of two images (NDVI and IR) is demonstrated in circles and rectangles illustrated in Fig. 4a and 4b. In the IR image (Fig. 4a) the burned area in the circle is not well identified as in the NDVI image. On the other hand, the cultivation area shown in the rectangle can be mistaken with the burned area in the NDVI image, whereas in the area in the IR image it is easily separated. Thus, the two images together provide a more accurate classification.

Fig. 4. Near infrared band - IR (a) and NDVI image (b)

Imaging processing is performed using the software MatLabR2007a, Multispec3.2 and

For better understanding, the proposed method is divided into five distinct phases. The first stage concerns data collection and image pre-processing. In the second phase there is NDVI calculation and the isolation of band 4 (near infrared - IR). The creation of routines that detect fires using two test images and that is based on a threshold T is performed in the third phase. The information on the existing biodiversity in the UUES and the preprocessing and data collection steps occur in the fourth phase. Finally, the fifth phase concerns analysis of the results based on the quantification of burned areas and climatic

All the nine images were geometrically corrected by means of image registration. The reference (georeferenced) image was also provided by INPE, at no cost, and the control points were collected in this image. Landsat TM images are first corrected by INPE, but with approximate coordinates, which simplifies the record of these scenes. For this reason, a firstdegree polynomial model and the interpolator "nearest neighbor", which maintains the original Digital Values DV (or Gray Levels), were used. Since the digital values define the different features in digital images, depending on the intensity of the changes of the digital

The spectral signature of the "burned" feature varies according to the fire status at the moment of image collection. If there is a fire, the digital values in the green (G) and red (R) bands are medium or close; and in the infrared band (IR), the DV are low. Though, when there is no fire, the DV in the IR band is null, spectral characteristics of the ash and coal. In the G band, the DV are low, being however larger than the values found in the R band. In this context, the pixels of the "burned" class can be classified by vegetation indexes, which

NDVI (Normalized Difference Vegetation Index) – index is selected because of its efficiency in the identification of burned areas. The method is widely recognized in the literature and

The isolation of the IR band (band 4 - TM) is a simple task, which is performed by ensuring the geometrical correction and cut out over the area of interest. There are no alterations in the original digital values in this band. Fig. 4 illustrates the image IR band and the resulting

The simultaneous use of two images (NDVI and IR) is demonstrated in circles and rectangles illustrated in Fig. 4a and 4b. In the IR image (Fig. 4a) the burned area in the circle is not well identified as in the NDVI image. On the other hand, the cultivation area shown in the rectangle can be mistaken with the burned area in the NDVI image, whereas in the area in the IR image it is easily separated. Thus, the two images together provide a more accurate

image of the NDVI calculation, both originated from the image shown in Fig. 3.

NDVI = (IR-R)/(IR+R) (1)

Envi4.2.

**5. Method** 

phenomena.

classification.

**5.1 Obtaining and pre-processing the data** 

values, it could be difficult to carry out fire analyses.

are based on the difference between the DV of the IR and R bands.

uses simple calculations. This index is calculated by equation 1

**5.2 Preparing the NDVI and IR images** 

#### **5.3 Identification of burned areas in digital images**

The areas corresponding to the "burned" features are identified in the IR and NDVI images (Fig. 4a and 4b). Due to the spectral characteristics of this feature (burned), the digital values in the NDVI image are negative and close to value -1. So, a threshold T is empirically established to isolate the burned areas from the other features in the image. All digital values of the NDVI image smaller or equal to T are labeled as belonging to the feature "burned", and the others (>T) as "unburned". The same procedure is used in IR imaging, but the magnitude of new threshold is T1.

The resulting images of threshold T and T1 applications are submitted to convolution with the morphological operator "opening", aiming to eliminate the noise in T and T1 operations. The opening of NDVI or IR images by operator B (structuring element) is obtained by the erosion of these images with B, followed again by dilation of the resulting image by B. The mathematical representation of the morphological operation with the NDVI image is given by equation 2.

$$\text{NDVI} \,\text{o}\,\,\text{B} = \begin{pmatrix} \text{NDVI} \,\text{o}\,\,\text{B} \end{pmatrix} \oplus \text{B} \tag{2}$$

The structuring element B can be defined in many ways: linear, circular, rectangular or diagonal. The entire process involved in the identification of burned areas is accomplished by routines developed in MatLab2007a. Finally, the burned areas obtained from the NDVI and IR images are added (arithmetic operation), to fill up spaces that were not visible with the use of other techniques.

#### **5.4 Classification of the Biome in UUES**

The classification of vegetation types affected by forest fires within the study area (UUES) is done by *in loco* visits. The present study was conducted by a group of researchers and members of ICMBio (Chico Mendes Institute for Biodiversity Conservation) responsible for monitoring the UUES. The group crossed the UUES, passing through a rural road towards the North and the South, and used also a side access road.

Some central points of the burned areas detected in the digital images are used as reference in the collection of information concerning the type of vegetation in this area. In each point,

Identification and Analysis of Burned Areas in Ecological Stations of Brazilian Cerrado 193

the third is MAM (March, April and May), and so on. This index is based on the limit of ±0.5 °C for the anomalous temperature at the sea surface (SST) in the Niño 3.4 region (5°N to 5°S latitude and 120°W 170°W longitude). Positive values of this index evince the occurrence of

In Fig. 6 the x-axis shows the year represented by the 12 quarters. Thus, year 1982 is represented by number 12, 1983 is represented by number 24, 1984 by 36, and so on. The

The constants used in this study are presented below. In the image registration procedure, the largest root mean square error (RMS) measured was 0.48 pixel. The values used for thresholds T and T1 in the identification of burned areas were 0.4 and 80, respectively. The structuring element B used in "opening" operation is the disc of radius 5. Essentially, opening removes small objects/noises (<5 pixels). The side effect is the elimination of the edges to round the objects. Although the burned areas may be partially (large areas) or totally (small areas) eliminated by the opening, this technique is very effective to eliminate

El Niño and negative values indicate the occurrence of La Niña.

noise. In Fig. 7 is shown an example in image IR-1989.

**6. Results and discussion** 

assessed years in the present study are shown in vertical lines in Fig. 6.

(a) (b)

eliminated with the application of that technique morphology.

Several points detected as burning alone (Fig. 7a) were removed by opening operation (Fig. 7b). The circles exposed in Fig. 7a show concentration of these points (noise), what were

The burned areas identified in the Landsat 5 TM images, concerning the nine years

Visually, it can be inferred that in 1985 there were few fires, which were more frequent across the borders of the UUES area. One possible explanation is that some areas of the UUES are inhabited, and burning is used for the practice of subsistence farming. In the following years (1987, 1989 and 1996) an abrupt expansion of the burned areas was observed, although with a small decline between 1996 and 1998. According to members of ICMBio, this increase was due to the advent of intensive agriculture and the absence of fire-

Fig. 7. Points eliminated by the opening operation.

analyzed, can be seen in Fig. 8.

the geodetic coordinates are recorded and the information of the vegetation physiognomy is collected in the neighborhood within a radius of approximately 200 meters. This information is stored as data points and then interpolated to other points on the area. Additionally, a photographic record is made at each point, e. g. in Fig. 5.

Fig. 5. Photographic record of local vegetation

Information on the vegetation is superimposed on burned areas, which are generated from the NDVI and IR, to measure and record biodiversity losses. The overlapping or crossing of this information is performed using routines developed in MatLab2007a.

#### **5.5 Behavior of climatic phenomena**

Finally, the results of all the experiments are assessed in relation to the climatic phenomena El-Niño and La-Niña, in order to determine whether or not these phenomena have influence on burnings. This analysis is done by checking the size of the burned area and the behavior of weather phenomena in the analyzed year. According to CPC (2010), the occurrence of the climatic phenomenon can be represented by indexes.

The quarterly values of the indices corresponding to the occurrence of El Niño and La Niña are graphically plotted (Fig. 6), in order to describe the behavior of these phenomena over time. The period represented on the graph varies between the years 1982 and 2010. Each year is punctuated by twelve quarters that overlap each other, for example, JFM (January, February and March) is the first quarter; the second is FMA (February, March and April);

Fig. 6. El Niño and La Niña based in limit of ±0.5 °C for the Oceanic Niño Index

the third is MAM (March, April and May), and so on. This index is based on the limit of ±0.5 °C for the anomalous temperature at the sea surface (SST) in the Niño 3.4 region (5°N to 5°S latitude and 120°W 170°W longitude). Positive values of this index evince the occurrence of El Niño and negative values indicate the occurrence of La Niña.

In Fig. 6 the x-axis shows the year represented by the 12 quarters. Thus, year 1982 is represented by number 12, 1983 is represented by number 24, 1984 by 36, and so on. The assessed years in the present study are shown in vertical lines in Fig. 6.
