Detection of Wildfire-Damaged Areas Using Kompsat-3 Image: A Case of the 2019 Unbong Mountain Fire in Busan, South Korea

Abstract

Forest fire is a critical disaster that causes massive destruction of forest ecosystem and economic loss. Hence, accurate estimation of the burned area is important for evaluation of the degree of damage and for preparing baseline data for recovery. Since most of the area size damaged by wildfires in Korea is less than 1 ha, it is necessary to use satellite or drone images with a resolution of less than 10m for detecting the damage area. This paper aims to detect wildfire-damaged area from a Kompsat-3 image using the indices such as NDVI (normalized difference vegetation index) and FBI (fire burn index) and to examine the classification characteristics according to the methods such as Otsu thresholding and ISODATA(iterative self-organizing data analysis technique). To mitigate the salt-and-pepper phenomenon of the pixel-based classification, a gaussian filter was applied to the images of NDVI and FBI. Otsu thresholding and ISODATA could distinguish the burned forest from normal forest appropriately, and the salt-and-pepper phenomenon at the boundaries of burned forest was reduced by the gaussian filter. The result from ISODATA with gaussian filter using NDVI was closest to the official record of damage area (56.9 ha) published by the Korea Forest Service. Unlike Otsu thresholding for binary classification,since the ISODATA categorizes the images into multiple classes such as(1)severely burned area, (2) moderately burned area, (3) mixture of burned and unburned areas, and (4) unburned area, the characteristics of the boundaries consisting of burned and normal forests can be better expressed. It is expected that our approach can be utilized for the high-resolution images obtained from other satellites and drones.

1. Introduction

Forests are important carbon absorption sources. By absorbing carbon dioxide in the atmosphere through photosynthesis and storing carbon in plants and soils, they can decrease the acceleration of global warming (KFS, 2009; Nayar, 2009). In addition, forests affect the microclimate, water circulation, and energy budget by releasing the water into the atmosphere through transpiration (Bonan, 2008; Davis et al., 2019). Although continuous efforts are required to preserve forests, unexpected wildfires have caused massive deforestation and economic loss. Reduced precipitation and dry weather lasting from winter peak in the spring, when forest fires in South Korea are the most frequent. The main causes of the forest fires in South Korea include farm field incineration, garbage incineration, and cigarette misfire (KFS, 2019).

Currently, the KFS (Korea Forest Service) gives efforts to estimate the area damaged by wildfires and damage amount immediately after the fire, which is then used as reference data for damage status checks, forest-fire prevention, and management (KFS, 2017). Wildfire-damaged area is calculated using data from field surveys, aerial photographs, digital maps, and GPS equipment. These approaches allow for a precise survey of forest fire damage; however, they are limited in that they are time consuming and expensive. To compensate for these limitations, satellite remote sensing can be applied in wildfire-damaged area detection techniques.

By continuously providing land surface information in the form of an image, satellites enable the detection and monitoring of forest fire damage without the need to inspect the site in person (Iverson et al., 1989; Holmgren and Thuresson, 1998). Currently, information regarding the global wildfire-damaged areas is provided with a resolution of 500 m by MODIS MCD64A1 (BurnedArea Product). However, because approximately 90% of forest fires occurring in South Korea are small, under 1 ha (KFS, 2019), the 500-m resolution is not sufficient for examination of wildfires in South Korea. Therefore, it is necessary to use satellite and drone images with a resolution of less than 10 m for South Korea.

Kompsat-3 (K3)is a low-earth-orbit satellite launched in 2012. It has a revisit cycle of 3 or 4 days, and the spatial resolution of the blue, green, red, and NIR(near infrared) bands is 2.8 m (Table 1). In terms of spatial and temporal resolution, it is reasonable to utilize K3 images to detect wildfire-damaged areas. Before calculating the full-scale damage area through field surveys, it may be effective to detect the damaged area using K3 images for rapid understanding of the disaster status.

Table 1. Spectral information of Kompsat-3

For satellite-based analysis of wildfire-damaged area, various indices such as NDVI (normalized difference vegetation index), NDWI (normalized difference water index), BAI(burn area index), and FBI (fire burn index), which were developed through spectral analysis of burned and normal forests, have been used (Won et al., 2014; Reszka and Fuentes, 2015; Lee et al., 2017; Park et al., 2019). In previous studies, to distinguish the burned forest in pixels, a binary classification method (e.g., Otsu thresholding) (Park et al., 2019; Bin et al., 2019) and unsupervised classification (e.g.,ISODATA(iterative self-organizing data analysis technique))(Won et al., 2014; Axel, 2018; Lasaponara and Tucci, 2019) were used. These pixel based classification approaches have the advantage of being universally applicable throughout a study area; however, they can face the problem of salt-and-pepper phenomenon (Roldán Zamarrón et al., 2006; Lee et al., 2015).To alleviate this problem, the GF (gaussian filter) can be applied through convolution and smoothing processes(Deng and Cahill, 1993;Yang and Qiu, 2015; Fawwaz et al., 2018).

This study aims to compare and evaluate the methodologies for detection of wildfire-damaged areas using high-resolution satellite images. To improve the accuracy of the detection of wildfire-damaged areas, NDVI and FBI were calculated using K3 images with 2.8-m resolution, and the results depending on the classification methods (Otsu thresholding vs. ISODATA) with and without using GF were also compared. Fig. 1 shows the flow chart of this study.

Fig. 1. Process for classification of burned area.

2. Data and Methods

1) Study area

Mountain Unbong islocated in Haeundae-gu, Busan Metropolitan City (Fig. 2). The wildfire occurred on April 2, 2019, and was extinguished on April 5, 2019. The cause of the fire was the incineration of garbage, and the area damaged was approximately 56.9 ha. At this time, the Busan area had a warning of dry weather, and the damage appears to have increased as strong winds blew (Korea Joongang Daily, 2019).

Fig. 2. Location of study area.

2) Data used

To calculate the burned area using NDVI and FBI, we used a K3 image with little cloud taken at 13:11 on June 5, 2019. For the K3 image, Level-1R data with radiometric correction and Level-1G data with additional geometric correction were provided in the GeoTIFF format (Lee et al., 2013). We used Level-1G data and conducted additional atmospheric correction using the FLAASH (fast line-of-sight atmospheric analysis of spectral hypercubes)tool. Also, to distinguish only forests, land-cover map provided by the Ministry of Environment was used.

3) Research method

(1) Wildfire damage index

Healthy vegetation mainly absorbs red wavelength of sunlight and reflects NIR light (Mahajan and Raj, 2016). NDVI (Eq. 1) shows the vitality of vegetation using these characteristics and has values between -1 and 1 (Escuin et al., 2008). NDVI values close to or less than 0 mostly represent bare soil and water area, and vegetation with higher vitality has an NDVI value close to 1, while the NDVI is lower in burned forests than in healthy forests (Escuin et al., 2008).

FBI is a normalized index using red versus green reflectance ratio (R/G) and NIR reflectance (Eq. 2). This index utilizes the property of the burned forests with a high ratio of red versus green reflectance and a low NIR reflectance. It has a range between -1 and 1, showing a high value in the burned forest (Lee et al., 2017).

\(\text {NDVI} = \frac {\text {NIR- Red}} {\text {NIR +Red}}\)       (1) 

\(\text {FBI} = \frac {\text {(Red/Green) -NIR}} {\text {(Red/Green) + NIR}}\)       (2)

We did not use NDWI and BAI because the NDWI requires a shortwave infrared band which is not included in the K3 image, and the empirical constants for BAI may not be suitable for Korean forests according to seasonality.

(2) Application of Gaussian filter

To mitigate the salt-and-pepper phenomenon that occurs during pixel-based classification, a GF was applied to the NDVI and FBI images. GF is used to perform convolution operations with each pixel using a weighted kernel generated through the Gaussian distribution function (Seddik, 2014; Fawwaz et al., 2018). Gaussian distribution is a bell-shaped function that is horizontally symmetrical. Its weight becomes smaller with distance from the center, and a larger standard deviation results in a greater smoothing effect (Seddik, 2014). In this study, the NDVI and FBI images were filtered using a GF with a standard deviation of 1.

(3) Wildfire-damaged area classification

The wildfire-damaged areas were classified by applying the Otsu and ISODATA methods to the four types of images: (1) NDVI w/o GF, (2) NDVI w/ GF, (3) FBI w/o GF, and (4) FBI w/ GF. The Otsu thresholding performs binary classification of the pixels in the image, by setting the optimal threshold to minimize the intra-class variance and maximize the inter-class variance (Otsu, 1979). ISODATA classifies the image into n classes through cluster analysis according to the spectral characteristics of pixels (Dhodhi et al., 1999).In the first step, the initial random average vector is assigned by the specified number of clusters, and the Euclidean distance is assigned to the nearest average by comparing the separate pixels to each cluster average. In the second step, a new average vector for each cluster is calculated, and the division and integration of each cluster is performed. Based on the new cluster average vector, each pixel is reallocated in a new cluster. The process of repeating the division and integration of these clusters continues until the class shows almost no change or until the maximum number of iterations set by the user is reached. In this study, the number of class for ISODATA was set to four, considering that the burned forest can be divided into several categories according to the severity of damage.

3. Results and Discussion

1) Wildfire-damaged area classification using Otsu thresholding

We first conducted a visual check on the RGB true color image and R-NIR-G false-color image, which showed contrasting color with brown as the burned forest and green as the normal forest(Fig. 3). Moreover, the NDVI and FBI images showed significant contrast between burned and normal forests (Fig. 4(a) and 4(c)). When the Otsu thresholding was applied to NDVI and FBI images, the threshold between the burned and normal forests was determined to be NDVI = 0.426 and FBI = -0.819. The burned forest was defined as the case when the NDVI was smaller than the threshold and the FBI was larger than the threshold. Fig. 5(a) and 5(c) show images of burned and normal forests classified by the Otsu thresholding. In the images, a slight salt-and-pepper phenomenon was observed in the boundary where the burned and the normal forests were separated, but the normal forest was mostly well distinguished from the burned forest.

Fig. 3. (a) RGB true-color image and (b) R-NIR-G false-color image of study area.

Fig. 4. Grayscale image of (a) NDVI w/o GF, (b) NDVI w/ GF, (c) FBI w/o GF, and (d) FBI w/ GF.

Fig. 5. Binary classification of burned and unburned areas using Otsu thresholding: (a) NDVI w/o GF, (b) NDVI w/ GF, (c) FBI w/o GF, and (d) FBI w/ GF.

NDVI w/ GF and FBI w/ GF images showed smoothed features overall compared with the results without using GF (Fig. 4(b) and 4(d)). The Otsu threshold was determined to be NDVI = 0.427 and FBI = -0.822, which was similar to the value before applying the GF. The results of the binary classification of burned and normal forests shows that the salt-and-pepper phenomenon was significantly reduced, and the boundaries of the burned forest are clearer (Fig. 5(b) and 5(d)).

2) Wildfire-damaged area classification using ISODATA

NDVI w/o GF, NDVI w/ GF, FBI w/o GF, and FBI w/ GF images were classified into four classes using ISODATA. The results were also visually compared with R-G-B and R-NIR-G images. Class 1, 2, and 3 were found to be the burned forest, and Class 4 was found to be the normal forest(Fig. 6). Class 1 occupied the largest area and was located in the center of the burned area. Class 2 was distributed in the area surrounding Class 1, and Class 3 was distributed on the boundary of the normal forest which is the outermost area of the burned area, showing a mixed characteristic between the burned and normal forest.

Fig. 6. Classification of multiple categories using ISODATA: (a) NDVI w/o GF, (b) NDVI w/ GF, (c) FBI w/o GF, and (d) FBI w/ GF.

When the GF was not used, the average NDVI was the lowest in Class 1 and increased sequentially for the Class 2, 3, and 4 (Fig. 7(a)). The average FBI was the highest in Class 1 and decreased sequentially for the Class 2, 3, and 4 (Fig. 7(c)). This is closely related to the spatial characteristics of Class 1, which is located in the innermost part of the burned area, and of Class 2 and Class 3 being increasingly distributed on the outskirts in the NDVI and FBI classification images, which indicates that the damage intensity of the forest is greater in the center of the burned area. The values of the NDVI and FBI of Class 3 consisted of the mixture of Class 2 (burned) and Class 4 (normal), which may include some misclassified pixels (Table 2).

Fig. 7. Distribution of NDVI and FBI for each class by ISODATA: (a) NDVI w/o GF, (b) NDVI w/ GF, (c) FBI w/o GF, and (d) FBI w/ GF.

Table 2. Minimum and maximum values for each class by ISODATA

It was possible to visually confirm that the salt-and-pepper phenomenon of Class 3 decreased when using GF (Fig. 6(b) and 6(d)). Also, the outliers and the variance of each class decreased in terms of the box plots of NDVI w/ GF and FBI w/ GF (Fig. 7(b) and 7(d)). Because of the smoothing effect of the Gaussian kernel, Otsu and ISODATA classification results Fig. 6. Classification of multiple categories using ISODATA: (a) NDVI w/o GF, (b) NDVI w/ GF, (c) FBI w/o GF, and (d) FBI w/ GF. showed a clearer distinction between the classes (Fig. 6(b) and 6(d)). To extract the wildfire-damaged areas from the ISODATA classification consisting of four classes, we examined two scenarios: (1) Class 1 and 2 were assumed to be the burned areas, and (2) Class 1, 2, and 3 were assumed to be the burned areas (Fig. 8). Class 3 contains about 30% the pixels of NDVI > 0.5, which implies that it is difficult to identify Class 3 as burned or unburned area. Although Class 3 may be classified as a normal forest to avoid salt-and-pepper phenomenon overall, it would be more reasonable to set Class 3 as a mixture of burned and normal forests because the outermost part of the burned area is partly included in Class 3. Therefore, it is possible to classify Class 1 as severely burned area, Class 2 as moderately burned area, Class 3 as mixture of burned and unburned areas, and Class 4 as unburned area (Fig. 6).

Fig. 8. Binary classification of burned area (class 1, 2, and 3) and unburned area (class 4) using ISODATA: (a) NDVI w/o GF, (b) NDVI w/ GF, (c) FBI w/o GF, and (d) FBI w/ GF.

Wildfire-damaged area recorded by KFS is 56.9 ha. In the ISODATA classification results, the values derived assuming class 1, 2, and 3 as burned areas were numerically closest to the official record (Table 3). Unlike the Otsu thresholding for binary classification, ISODATA classifies into several classes. As a result, the identification of the area where burned and unburned pixels coexist was possible by the ISODATA. Unlike buildings and roads, areas damaged by forest fire have no clear boundaries. Therefore, the burned area calculated including the ambiguous areas such as Class 3 is considered to be more appropriate. Our approach can be utilized for the high-resolution images obtained from drones. More detailed information about burned areas with a submeter resolution can be obtained by using drones; however, battery duration still remains a challenging task for examination of a wide area.

Table 3. Burned area by classification methods (ha)

4. Conclusion

In this study, we conducted the detection of wildfire-damaged areas from a K3 image using the indices such as NDVI and FBI with the classification methods such as Otsu thresholding and ISODATA. For a case study for Mountain Unbong in Busan, we compared the results from the two methods and the mitigation of salt-and-pepper phenomenon depending on the use of GF. The contrast between burned and unburned areas was clear in the NDVI and FBI images, which means that NDVI and FBI are suitable for detection of wildfire-damaged areas. Both Otsu thresholding and ISODATA could perform the classification well, and the salt-and-pepper phenomenon was reduced by the use of GF. Unlike the Otsu thresholding, ISODATAcould express the mixture of burned and unburned areas as well as the severely burned area and moderately burned area, which produced the closest estimation to the official record provided by KFS. For a more reasonable assessment of the methodologies, more wildfire cases in Korea should be tested with the consideration of seasonal changes. Also, terrain effects like shadow should be considered for a more reliable result. Moreover, development of an optimal methodology for wildfire-damaged areas using the next-generation low-earth-orbit satellites will be necessary for a future work.