Abstract: The study of burned areas is used as a subsidy for fire control and monitoring in the protected areas. In face of the challenges of the spectral signature characterization of burned areas, this study aimed to apply the object-oriented classification method and to evaluate the performance of spectral indices subsets for mapping burned areas in the Atlantic Forest. For that, we performed a bitemporal analysis between 2014 and 2016, considering the difference of each spectral indices among two LANDSAT 8 images: pre-and post-fire. The object-oriented classification was performed automatically by segmentation, supervised classification and optimization algorithms in the GIS environment. The “weak” burn severity class was the most expressive, with 13.65% of the mapped area, while the “severe” burn severity class occupied 0.3%. The burned areas presented an increase of reflectance in the red and shortwave infrared bands and a decrease in the near infrared band. The ΔNBR was the best discriminator of burned area and the ΔNBR, ΔNBR2, ΔNDMI, ΔSAVI, ΔNDVI, ΔGEMI and ΔMSAVI set presented the highest separation threshold. The validation of the classification by the Kappa agreement coefficient obtained a good outcome (0.72). The selection of the variables showed efficiency in determining the spectral indices’ subset with the best performance for detecting the classes of burned areas, improving the classification accuracy and reliability. The segmentation was also important for the effectiveness of the object-oriented classification, being directly influenced by the image spatial resolution.
Keywords: Image classificationImage classification,Image segmentationImage segmentation,Band ratioBand ratio,Variables selectionVariables selection,Forest fireForest fire.
Articles
Bitemporal analysis of burned areas in the Atlantic Forest
Iorrana Figueiredo Sacramento iorranafigueiredos@gmail.com
Roberto Ferreira Machado Michel roberto@michel.com
Rafael Gomes Siqueira rafael.geo.siqueira@gmail.com
Received: 24 March 2020
Accepted: 16 June 2020
The Atlantic Forest (AF) is a biome historically threatened by the anthropic pressure and it is considered a worldwide hotspot that presents altered and fragmented areas (PINTO et al., 2006). Between 1986 and 2001, although the AF has increased 5% of its absolute cover in the South of Bahia (BA), Brazil, there was an intense fragmentation of the landscape, with an increase in the number of fragments smaller than one hundred hectares (SAVE BRASIL; IESC; BIRDLIFE INTERNATIONAL, 2009). Studies in the Biological Reserve of Una - BA (REBIO of Una) showed a great quantity of cabrucas (agroforestry systems where the cocoa trees are planted under the forest shadows) and secondary forests, which contribute to decrease the isolation and maintain the populations of animals and plants in the region. This fact emphasizes the importance of maintaining contiguous areas of protected areas.
The burned areas represent a recurrent threat for the protected areas (SANTOS; SOARES; BATISTA, 2006; 2018). The use of fire for soil management has a millenary nature (BENTO-GONÇALVES et al, 2012; CHUVIECO, 1997; MISTRY; BIZERRIL, 2011). The traditional management is the main agricultural strategy in the Wildlife Refuge of Una - BA (REVIS of Una) (SOLLBERG; SCHIAVETTI; MORAES, 2014) and, despite this technique embraces conservationists’ practices, it also uses slash-and-burn agriculture. The fire consequences affect the forests’ composition and structure, the animal biodiversity, the human health and the atmospheric balance, both on a local and regional scale (COCHRANE, 2003; DODONOV et al., 2019; FREITAS et al., 2005; MENEZES; CAZETTA; DODONOV, 2019).
The creation of fire prevention and monitoring politics has become possible through the application of remote sensing techniques, once they allow extracting information quickly and efficiently, which are used for burned areas identification, control and supervision (CHUVIECO, 1996; DOS SANTOS et al., 2018; FREITAS et al., 2005; PEREIRA, 2009; SANTOS et al., 2020). The bitemporal analysis of satellite images has been widely used for detecting and monitoring environmental changes, mainly for the estimation of burned areas and the detection of fire severity(CARDOZO et al., 2011; CARVAJAL-RAMIREZ et al., 2019; ESCUÍN; NAVARRO; FERNÁNDEZ, 2008; SANTOS et al. 2020). The object-oriented classification is a valuable method for considering not just the spectral information, but also the spatial relationships as shape and context. As consequence, the outcome is a more sophisticated classification and closer to human interpretation, without the common “salt and pepper” effect of the pixel-by-pixel method (McDERMID, et al., 2003; WONG et al., 2003).
However, burned areas detection using remote sensing is limited due to the difficulty to discriminate fire from possible imaging failures, that can restrict the detection (COCHRANE, 2003). Besides, the confusion tendency between burned areas and other land uses has led to the need for a better characterization of the spectral signature of burned areas, which has been possible through spectral indices. These indices help discriminating land surface targets, through arithmetic calculation with spectral bands (CARVAJAL-RAMIREZ et al., 2019; DOS SANTOS et al., 2018; LEITE et al., 2017; SANTOS et al., 2020). Therefore, the main goals of this research were to apply the object-oriented classification method and to evaluate the best performance of sets of spectral indices’ for mapping burned areas in the AF of South of Bahia.
The study area (709.7 km²) comprises the protected areas of integral protection: REBIO of Una (187 km²), REVIS of Una (232 km²) and their surrounding areas (a radius of 3 km). These units were created in 1980 and 2007, respectively (BRASIL, 1980; 2007). The study area is in the municipality of Una, South of Bahia’s region (Figure 1). The estimated population of Una is 19.000 inhabitants and its demographic density is 16.86 inhabitants per km² (IBGE, 2019). Its economy is mainly based on agriculture and extractivism, highlighting the cocoa production (cabrucas), rubber agriculture, coffee, manioc, coconut, açaí palm and banana (IBGE, 2017).
The study area integrates the Serra das Lontras Complex, formed by three connected mountains that are part of the Serras e Maciços Pré-Litorâneos geomorphological unit (SAVE BRASIL; IESB; BIRDLIFE INTERNATIONAL, 2009). The region is characterized by well-distributed rains over the year, with annual precipitation around 1,400 mm and annual mean temperature of 24.8 ºC (Figure 2). The air relative humidity was 84% in 2014 (INMET, 2020). These conditions provide the formation of deep, well-drained and porous soils - Yellow and Yellow-Red Latosols - originated from crystalline rocks (SAVE BRASIL; IESB; BIRDLIFE INTERNATIONAL, 2009). The Dense Ombrophilous Forest currently corresponds to 81% of the study area (Figure 1) and had increased about 1% since 1988 (MAPBIOMAS, 2020), evidencing the importance of the protected areas for the conservation of the AF remaining fragments.
We used the 30 meters spectral resolution bands of the Operational Land Imager Sensor - OLI, LANDSAT 8 (path/row 215/71) (Table 1). These products were obtained post-processed, with geometric and surface reflectance corrections - atmospheric absorption or scattering correction (BARSI et al., 2014; VERMOTE et al., 2016). The 15 meters panchromatic band was only used for the segmentation step to enhance the detailing. The pre-fire image was from August 7th, 2014 and the post-fire image was from April 6th, 2016. We selected images from different periods due to the lower cloud cover and to the absence of a marked dried period in the region (Figure 2), which ensured similar spectral response of the vegetation over the year.
We obtained the hot spots data from the Banco de Dados de Queimadas (BDQ) of the Instituto Nacional de Pesquisas Espaciais (INPE, 2018), the Brazilian spatial institute. These data were generated by polar satellite images NOAA and EOS, and geostationary satellite images GOES and METESAT. Between 2014 and 2016, 224 hot spots were identified in the study area. These information were used as evidence of fire occurrence and were applied in the classification’s validation step.
We performed the processing in eCognition 9 software. For the classification, the clouds, shadows, and water bodies from the pre-and post-fire images were removed. From the post-fire image, we created five classes of change detection: 1 - severe, 2 - moderate, 3 - weak (burn severity classes), 4 - no alteration class, and 5 - regeneration class (change occurred by biomass increasing). Burn severity classes followed the criteria: the sever class corresponded to the burned areas’ epicenter, the “weak” class is the edge zones, and the “moderate” class is the transition between “severe” and “weak” classes. For each class, we collected 15 training samples, totalizing 75 samples.
All bands from the post-fire image were used in the segmentation step. We applied the Multiresolution Segmentation algorithm (BAATZ; SHÄPE, 2000). This algorithm initiates the segmentation with each image pixel representing one object, which can be grouped to a neighbor pixel according to the scale and homogeneity parameters. The first parameter defines the maximum heterogeneity allowed in an object and it influences its size. The second refers to a color and geometry relationship of the objects aiming to create objects as less heterogeneous as possible (DEFINIENS, 2012). We made tests to define these parameters and the values of scale 80, shape 0.2 and compactness 0.5 showed more suitability for the study area.
Then, we calculated the following spectral indices through the bands ratio, for the pre-and post-fire images: Normalized Difference Vegetation Index - NDVI (Equation 1) (ROUSE et al., 1973); Enhanced Vegetation Index- EVI (Equation 2) (JENSEN, 2009); Soil Adjusted Vegetation Index - SAVI (Equation 3) (HUETE, 1988); Modified Soil Adjusted Vegetation Index - MSAVI (Equation 4) (QI et al., 1994); Normalized Difference Moisture Index - NDMI (Equation 5) (JESEN, 2009); Normalized Burnt Ratio - NBR (Equation 6) (KEY; BENSON, 1999); Normalized Burnt Ratio 2- NBR2 (Equation 7) (MARTÍN; GÓMEZ; CHUVIECO, 2005); Burned Area Index - BAI (Equation 8) (CHUVIECO; MARTÍN; PALACIOS, 2002); Global Environmental Monitoring Index - GEMI (Equation 9a e 9b) (PINTY; VERSTRAETE, 1992).
At which:
Then, we calculated the delta-images, that correspond to the difference between the pre-fire and post-fire indices images (Δindices = pre-fire index - post-fire index). As a result, the delta-images areas where the vegetation suppressed obtained values > 0; where there was regeneration, the values were < 0; and where there was no change, the values were close to 0 (CARDOZO et al., 2011; PEREIRA, 2009).
Next, we used the Feature Space Optimization tool to define which delta-images were more suitable for the classification. This tool ranks the variables (delta-images) by importance order and creates a number of different size subgroups (dimensions), proportional to the number of the cumulative variables (dim 1 = one variable; dim 2 = two variables, dim 3 = three variables… dim n = n variables). For each dimension, the algorithm calculates the Euclidean distance (from 0 to 1) between the objects of each mapped class, in order to find the variable combination that presents the greater distance and, consequently, the greater classes separation threshold (TRIMBLE, 2014).
From this, we performed the supervised classification through the Nearest Neighbor algorithm. This algorithm works considering the distance between the object to be classified and the nearest training sample, in other words, as the closer is the object of a sample, the greater the relationship between them (GABRIEL, 2013), and then it is grouped to the sample’s class. The Nearest Neighbor was the algorithm with the best performance among the object-oriented classifiers tested by Zanotta et al. (2010).
For the validation step, we applied the Kappa index considered, in many works, as appropriate to measure a classification accuracy (GABRIEL, 2013). This index was applied through the Assessment Accuracy tool and for that, we collected 65 samples based on the INPE’s hot spots. According to Landis e Koch (1977), the Kappa index values range from 0 to 1, being 0 - 0.2 = bad; 0.2 - 0.4 = reasonable; 0.4 - 0.6 = good; 0.6 - 0.8 = very good; and 0.8 - 1.0 = excellent. The whole processing is briefly presented below (Figure 3).
Figure 4a shows the average reflectance values of samples of burned areas for the pre-and post-fire spectral bands. There was an increase of the reflectance at the visible region (RGB) and at the mid-infrared (SWIR1 and SWIR2), with higher alteration in the visible at the red (R) band. In the pre-fire image, with greater biomass condition, the reflectance at the R was less than the green (G) band, and the inverted situation happened in the post-fire image.
On the other hand, the near-infrared band (NIR) had a significant reflectance decrease (MARTÍN; CHUVIECO; 1998). This led to a decrease in the difference between the NIR and R bands, and between the NIR and SWIR bands, which affected directly the spectral indices behavior. Figure 4b shows the reflectance pixel values of burned areas for the R, NIR and SWIR2 bands. The reflectance values of the burned areas at the R and SWIR2 bands were higher in the post-fire image, while the values of the NIR band were higher in the pre-fire image.
The reflectance increase in the visible region, as well as the reflectance increase at the R band compared with the G band, is explained by the decrease of the radiation absorption of the photosynthetic pigments, due to the biomass loss caused by the burning (PONZONI; SHIMABUKURO, 2007). Despite this difference, the visible region is not indicated for burned areas discrimination, due to its low reflectance and to the atmospheric scattering influence, mainly over dark surfaces, as burned areas, which decreases the contrast between the surface targets (PEREIRA et al., 1999).
The NIR reflectance decrease (around 30%) can be explained by the plant cell structure destruction during the burning. Due to the great reflectance difference of the NIR in the pre-and post-fire images, this band was highlighted as the best one for burned areas discrimination (PEREIRA et al., 1999). The significant increase of the reflectance values for the SWIR1 and SWIR2 bands (75.4% and 94.7%, respectively) is due to the decrease of the leaf water content (MARTÍN; CHUVIECO, 1998). The mid-infrared region was recognized as effective to the distinction of burned areas, due to its high reflectance and to the minimum atmospheric scattering influence (PEREIRA et al., 1999).
The Feature Space Optimization algorithm was efficient in the selection of the more suitable delta-images for classification (Figure 5 and Table 2). Considering just one delta-image in the analysis (dimension 1), the ΔNBR was the one with the greater capacity of change detection classes discrimination. However, the biggest separation threshold (0.564) was found when considering seven from nine calculated indices (dimension 7). From dimension 7, the high number of delta-images induced redundancies, which decreased the classes separation threshold. The Feature Space Optimization algorithm efficiency to enhance image classification performance was already highlighted in earlier studies (DURRIEU et al., 2007; LEDUC, 2004).
The NBR and NBR2 burned area indices were the most significant in this study and the ones that also discriminate better the burned areas according to earlier studies (CARDOZO et al., 2011; ESCUÍN; NAVARRO; FERNÁNDEZ, 2008; PEREIRA, et al., 2015; ROSAN; ALCÂNTARA, 2015). This is because the bands used in their calculation (NIR and SWIR2) are the ones that suffered the most significant post-fire spectral alterations. The ΔNDMI was the third-best delta-image for discriminating the change detection classes of burned areas. The NDMI humidity index also showed greater potential for burning detection than the vegetation indices according to Wilson and Sader (2002), which is due to the changes of soil and plant humidity regime after burning.
Among the vegetation indices, the SAVI had the greatest capacity of burning detection and it also presented moderate performance according to Pereira et al. (2015). The MSAVI is similar to SAVI, although the last one uses a soil effect correction factor manually adjusted according to the environmental conditions of each area and the first one uses an auto-adjusted factor (QI et al., 1994), whose hardness can induce to poor results comparing to SAVI, as verified in this research. In this study, the NDVI did not present great efficacy for burned areas detection, although it is widely applied (CARDOZO et al., 2011; ESCUÍN; NAVARRO; FERNÁNDEZ, 2008; PEREIRA, 1999; PEREIRA, et al., 2015). The GEMI proved to be better than NDVI for burned areas detection since it is less affected by atmospheric and soil influence (PEREIRA, 1999). Herein, the performance differences between the NDVI and GEMI delta-images were not significant, and both of them showed a moderate performance. The BAI did not contribute to the class separability enhance, it is considered a good burned area detection index though (CHUVIECO; MARTÍN, PALACIOS, 2002; DEMPWOLF, et al., 2007). As the BAI, the EVI did not contribute to the increase of the burned areas' differentiation and presented a low separability according to Pereira et al. (2015).
According to the Euclidian distance matrix between the change detection classes (Table 3), the “no alteration” and “regeneration” classes presented the lowest separation value (0.56), once they correspond to vegetated areas and therefore, they had similar spectral responses. The “severe” burned class presented the highest separability values with the two previous classes (28.6 and 37.7, respectively) due to the great biomass loss associated with the burning. Among the burning severity classes (classes 1, 2 and 3), the “severe” and “moderate” classes showed the lowest separation value (0.9) and the “severe” and “weak” classes showed the biggest one (4.7). Comparing the distances between the burned area classes and the others, the “weak” and “no alteration” classes presented the smallest separation value (5.3).
The “weak” severity class corresponds mainly to the burned areas edge, where the spectral difference between the pre-and-post-fire images was smaller, and the “severe” class corresponds to the burning epicenter, where the changes induced by the fire were bigger. The proximity between the distance values of the “weak” class with the “severe” and “no alteration” classes indicates a transitional condition of the “weak” class in the study area.
According to the final classification, around 74% of the total area did not pass through significant changes in the vegetation cover during the analyzed period (Figure 6 and Table 4). The burned area classes represent 16.1% of the mapped area and they were more pronounced in the northeast sector, which corresponds to the REVIS of Una and its surroundings (Figure 6), a sector where the land use is allowed and more intense. From this percentage, 13.65% corresponds to the “weak” class, while 0.3% corresponds to the “severe” class (Table 4).
Nevertheless, we also observed an overestimation of the “weak” class. The overestimation could have occurred for three reasons: occurrence of areas where the change was less noticed, with positive values of delta-images next to 0, as in the coastal areas; areas around clouds, whose influence resulted in positive values next to 1; or where there was a confusion generated by the presence of misclassified pixels inside some objects. This can be associated with the low spatial resolution of the LANDSAT8 images, that interfered in the segmentation process, once unburned pixels were included inside some burned area classes, mainly the “weak” class.
From the total mapped burned areas (95.23 km²), 45% were identified in the surrounding area; 34% in the REVIS of Una; and 21% in the REBIO of Una, indicating the best effectivity of the last one in preventing fires. The classification Kappa value was 0.72, corresponding to a very good classification according to Landis e Koch (1977).
The use of delta-images from pre-and post-fire spectral indices showed efficiency to the land cover change detection through burnings in the Atlantic Forest in the South of Bahia. With the variable’s selection using the Feature Space Optimization algorithm, we identified the most suitable spectral indices’ subset to enhance the object-oriented classification accuracy. The NBR delta-image showed the best performance, while the seven spectral indices’ subset presented the greater separation threshold, embracing a combination of burning, humidity and vegetation indices.
Our results showed that the integrated use of different spectral indices potentializes the classifier performance when compared to the use of the indices particularly. However, we observed that the performance increase is not directly proportional to the variables increase, once the use of more than seven indices increased the information’s confusion and redundancy, which contributed to reduce the separability level between classes.
The object-oriented classification also showed strong dependency of the segmentation step, once the classification is directly applied over the segmentation’s created objects, which shall involve as precisely as possible the pixels correspondent to the same surface target. The medium to low spatial resolution of the LANDSAT8 images induced a certain spectral confusion among the evaluated classes. An alternative to obtain more homogeneous objects in limited spatial resolution images would be the use of the spectral mixing method, that permits the identification of the different target's influence (vegetation, soil, shadow, e.g.) in a burned area pixel.
