Mapping of Post-Wildfire Burned Area Using KOMPSAT-3A and Sentinel-2 Imagery: The Case of Sokcho Wildfire, Korea

Abstract

On April 4, 2019, a forest fire started in Goseong County and lasted for three days, burning the neighboring areas of Sokcho. The strong winds moved the blaze from one region to another region and declared the worst wildfire in South Korea in years. More than 1,880 facilities, including 400 homes, were burnt down. The fire burned a total area of 529 hectares (1,307 acres), which involved 13,000 rescuers and 16,500 military troops to control the fire occurrence. Thousands of people were evacuated, and two people are dead. This study generated post-wildfire maps to provide necessary data for evacuation and mitigation planning to respond to this destructive wildfire, also prevent further damage and restore the area affected by the wildfire. This study used KOMPSAT-3A and Sentinel-2 imagery to map the post-wildfire condition. The SVM showed higher accuracy (overall accuracy 95.29%) compared with ANN (overall accuracy of 94.61%) for the KOMPSAT-3A. Moreover, for Sentinel-2, the SVM attained a higher accuracy (overall accuracy of 91.52%) than the ANN algorithm (overall accuracy 90.11%). In total, four post-wildfire burned area maps were generated; these results can be used to assess the area affected by the Sokcho wildfire and wildfire mitigation planning in the future.

1.Introduction

On April 4, 2019, a forest fire started in Goseong County and lasted for three days, burning the neighboring areas of Sokcho, Gangneung, and Inje (Nam, 2019). The strong winds moved the blaze from one region to another region and declared the worst wildfire in South Korea in years. Thousands of people are evacuated, and two people are dead (Kim and Kim, 2019). According to the fire control officials in Gangwon on National Public Radio (NPR) report, the fire burned a total area of 529 hectares (1, 307 acres), which involved 13, 000 rescuers and 16, 500 military troops to control the fire occurrence (Chappell, 2019). More than 1, 880 facilities, including 400 homes, were burnt down, and losses from damaged agricultural facilities alone were estimated at 5.2 billion won ($4.6 million) (Nam, 2019).

As a natural phenomenon, forest fires occurred in many ecosystems across the world that are responsible for burning about 350 million hectares of forested land per annum on an average-basis (FAO: Food and Agriculture Organisation, 2007). However, forest fire has positive and negative consequences on the ecosystem and people in many ways (Martell, 2011). The advantages from forest fire such as enhancing local species richness and diversity, regulating fuel accumulations, fungi, and microorganisms eradication, control the disease and insect, also releasing the mineral soil exposure and nutrient (Ruokolainen and Salo, 2009), still forcing a high level of interest in evaluating the forest fire in terms of climate change concerns. Since there are several drawbacks from forest fire too such as the release of CO2 into the atmosphere (Page et al., 2002), threatened health problem (e.g., inhalation of toxic gases from smoke worsen the heart and lungs diseases, sore eyes, tears, cough, and breath) (Stefanidou et al., 2008). Also, making the forest function disappear or deteriorate, increasing the risk of disasters such as soil leakage and landslides. If heavy rain pours into the area affected by a forest fire, it immediately flows into the valley, leading to secondary and tertiary disasters accompanied by changes in ecosystems and surface areas along with landslides or river flooding. Therefore, it is very important not only quickly extinguish a forest fire to reduce its damage but also to prevent and restore further damage to the area affected by the forest fire.

Previous studies mentioned that a remote sensing- based method is a powerful tool for monitoring the forest fire (Chowdhury and Hassan, 2015) and potentially in mapping burn severity (Vega-García and Chuvieco, 2006). Veraverbeke et al. (2010) used Landsat Thematic Mapper (TM) images corrected for geometric, radiometric, atmospheric, and topographic aspects to assess burn severity in the Peloponnese Peninsula in Greece. Another study by Wu et al. (2015) used Landsat-8 and WorldView-2 satellites to estimate vegetation burn severity from the Creek Fire in San Carlos. Syifa et al. (2020) used Landsat and Sentinel- 2 to map the post-wildfire burned area using Support Vector Machine (SVM) and SVM-ICA (Imperialist Competitive Algorithm) from Camp Fire in California, USA. Lee and Jeong (2019) used Korea Multi-Purpose Satellite (KOMPSAT) -3A satellite to classify forest fire severity using probability density function from wildfire in Gangneung, Korea. The methods used by those authors involved the application of two temporal images to differentiate and to ratio between pre-fire and post-fire images, principal component analysis for pre- fire and post-fire imagery, and machine learning; all of the methods demonstrated potential value for providing quantitative fire severity maps.

Along with remote sensing, image classification techniques from machine learning and other various approaches have been utilized in many areas of hazard assessment and disaster risk management, especially in mapping burned areas. For example, fuzzy logic was applied for Moderate Resolution Imaging Spectroradiometer (MODIS) data (Keramitsoglou et al., 2013), an artificial neural network (ANN) and SVM were applied to Landsat TM data (Zare et al., 2013), a hybrid algorithm was performed to Landsat and Sentinel-2 (Syifa et al., 2020).

Despite the large number of studies that have been published regarding burned area mapping using various machine learning, there is a lack of research using KOMPSAT-3A. By combining high-resolution multispectral data and machine learning, this study aims to map the post-wildfire in Sokcho using high- resolution KOMPSAT-3A and Sentinel-2 imagery with the same acquisition date and band combination. Four post-wildfire maps from two satellite imageries were generated using ANN and SVM classifiers and compared the mapping results to improve fire mitigation, management, and evaluation effectiveness.

2. Study Area

The Sokcho wildfire occurred from Goseong County and lasted for three days, burning the neighboring areas of north Sokcho. Fig 1(b) shows the area of wildfire in Sokcho (38°13′00″ N, 128°32′00″ E), east coast of Korea (Fig. 1(a)). The north side of Sokcho is a border with Goseong County, consist of agriculture and forest. The topography of the burned area is varied, ranging from gentle to very steep, with an elevation range of approximately from 0 to 110 meters above sea level (Digital Elevation Model from Shuttle Radar Topography Mission). On the day of the wildfire, the weather was dry with strong southerly winds, leading to extreme and fast fire behavior.

Fig. 1. (a) Sokcho in the eastern Korean Peninsula, (b) KOMPSAT-3A scene in natural color composite image taken on April 20, 2019.

The KOMPSAT series of satellites is a multi- purposed satellite system developed by Korea Aerospace Research Institute (KARI). The recent satellite of the KOMPSAT series, KOMPSAT-3A, is a sister spacecraft of KOMPSAT-3 (Arirang-3) and is Korea’s first earth observation/infrared satellite with two imaging systems on board. It is equipped with two imaging payloads; the Advanced Earth Imaging Sensor System A (AEISS-A), which has 0.55 m Ground Sample Distance (GSD) for panoramic image, 2.2 m GSD for multispectral image, and Infrared Imaging payload. A KOMPSAT pictures the earth’s surface from the outside atmosphere at an altitude of approximately 528 km and regularly revisits within 1.4 days (Jeon et al., 2016).

Sentinel-2, launched on June 23, 2015, consists of two satellites, namely Sentinel-2A and Sentinel-2B. It provides global optical imagery with 13 spectral bands, ranging in spatial resolution from 10 m to 60 m (Wanget al., 2017). In particular, Sentinel-2 has 10 m and 20m reflective wavelength bands that are well suited for burned area mapping (Huang et al., 2016). The satellite regularly revisits within ten days, and the data are freely accessible; therefore, Sentinel-2 data have great utility for mapping the approximate day of burning (Roy et al., 2019).

The high-resolution multispectral data used in this study was taken by KOMPSAT-3A and Sentinel-2 on April 20, 2019. The provided KOMPSAT-3A data consisted of Blue (450-520 nm), Green (520-600 nm), Red (630-690 nm), Near InfraRed (760-900 nm) were combined to produce a false-color view, as shown in Fig. 2(a). Also, the Sentinel-2 bands 3 (560 nm), 4 (665 nm), and 8 (864 nm) were combined to produce a false- color view, as shown in Fig. 2(b). The image quality of the study area from KOMPSAT-3A and Sentinel-2 were good and cloud-free. For better classification inputs, we derived Normalized Difference Vegetation Index (NDVI) (Rouse et al., 1973) using equation 1.

\(N D V I=\frac{N I R-R e d}{N I R+R e d}\)       (1)

 

 

Fig. 2. (a) Post-wildfire event captured by KOMPSAT-3A on April 20, 2019, and (b) Post-wildfire by Sentinel-2 on April 20, 2019, in false color combination.

All original bands and derived bands were composited to form a new multispectral image with five bands. Table 1 shows our classification scheme and category definitions. We classified burned area mapping into five classes using satellite images (Table 1).

 Table 1. Classification scheme and category definition for the study area

3. Method

1) Atmospheric Correction

Atmospheric correction of the top of atmosphere (TOA) reflectance to surface reflectance is required to implement the burned area mapping algorithm reliably (Kaufman and Tanre, 1994). The acquired Sentinel-2 was Level-1C, and KOMPSAT-3A was Level-1G products includes TOA reflectance. Sentinel-2 image was atmospherically corrected using algorithm ‘sen2cor’ v2.2.1 issued as part of the standard Sentinel-2 Toolbox (Roteta et al., 2019). This algorithm generated Top of Canopy (TOC) reflectance images and several products such as aerosol optical thickness, water vapor, and scene classification maps.

KOMPSAT-3A image was atmospherically corrected using algorithm supported in Orfeo Toolbox (OTB), a remote sensing open-source tool (Grizonnet et al., 2017; Lee and Kim, 2019). Equations (2), (3), and (4) below show the formulas for obtaining TOA reflectance and TOC reflectance according to atmospheric correction for KOMPSAT-3A image information.

\(L_{\lambda}=G a i n \times(D N)+O f f s e t\)       (2)

Here, L_λis the band-averaged spectral radiance at wavelength λ, and Gainand Offsetrepresent the gain and deviation of the sensor, respectively. Here, the data presented in KOMPSAT-3A Image Data Manual v1.5 (2019) can be applied for the sensor radiation value for each wavelength band of KOMPSAT-3A.

\(\rho_{\lambda T O A}=\frac{\pi L_{\lambda} d^{2}}{\left(E S U N_{\lambda}\right) \cos \theta_{s}}\)       (3)

Where ρ_λTOArepresents the TOA reflection for the satellite image band wavelength λ, ESUN_λ is band- dependent mean solar exo-atmospheric irradiance (Wm^-2μm^-1), θ_s is solar zenith angle, and dis the distance between the earth and the sun on the date of imaging.

\(\rho_{T O C}=\frac{\frac{\rho_{T O A}-\rho_{a t m}}{T\left(\mu_{s}\right) T\left(\mu_{v}\right) t_{g}}}{1+S \frac{\rho_{T O A}-\rho_{a t m}}{T\left(\mu_{s}\right) T\left(\mu_{v}\right) t_{g}}}\)       (4)

Where ρ_TOCis the TOC reflectance on the surface of the diffusely reflective surface assuming a homogeneous environment, ρ_TOAis the TOA reflectance, ρ_atmis the intrinsic atmospheric reflectance, Sis the atmospheric correction factor, T(μ_s), and T(μ_ν) are downward transmittance and upward transmittance, t_g represents the spherical albedo of the atmosphere. Aerosol measurement properties, such as AERONET data, can be used to generate TOC reflectance images. However, the data from the nearest AERONET site in Gangneung is absent. Therefore, we used the aerosol model for urban area.

2) Normalized Burn Ratio (NBR)

Sentinel-2 measures the earth’s reflected radiance over 13 spectral bands. These spectral bands span from visible and near-infrared (NIR) to short wave infrared (SWIR) at different resolutions (Navarro et al., 2017). Therefore, we can generate a Normalized Burn Ratio (NBR) index map using equation 5.

\(N B R=\frac{N I R-S W I R}{N I R+S W I R}\)       (5)

Healthy vegetation has very high near-infrared reflectance and low reflectance in the shortwave infrared portion of the spectrum. On the other hand, burned areas have relatively low reflectance in the near- infrared and high reflectance in the shortwave infrared band. A high NBR value generally indicates healthy vegetation, while a low value indicates recently burned areas (Navarro et al., 2017). NBR index map was generated from the Sentinel-2 image, as shown in Fig. 3, and used as a reference map to generate burned area class of training and test data (Lee and Jeong, 2019). We generate train and test data for the burned area class if the NBR index value is lower than 0.

Fig. 3. Normalized Burned Ratio (NBR) index map from Sentinel-2 on April 20, 2019.

3) Artificial Neural Network (ANN)

We used a feed-forward ANN, the most widely used classifiers, featuring one input layer, at least one hidden layer, and one output layer. Each layer is formed of non- linear processing units, termed neurons, and the connections between neurons in successive layers are weighted (Xiu and Liu, 2003). Only forward connections are allowed (thus from the input to the hidden layer, or from a hidden layer to a subsequent hidden or output layer; Li et al., 2014). Non-linear processing is achieved by applying an activation function to the summed inputs to each unit. Backpropagation is a gradient-descent algorithm minimizing the error between the outputs of training input/output pairs and the actual network outputs (Bishop, 1995). Therefore, a set of input/output pairs is repeatedly presented to the network, and the error is propagated from the output back to the input layer. The weights of the backward paths are updated using an update rule and the learning rate (Collobert and Weston, 2008). ANNs are not uniquely specified by the characteristics of their processing units or the training or learning rules selected. The network topology (i.e., the numbers of hidden layers and units, and their interconnections) also influence classifier performance. Here, we used the network architecture and training patterns suggested by (Kavzoglu and Colkesen, 2009). To form an ANN using ENVI software, we employed a logistic activation method. The training threshold contribution and training momentum field were set to 0.9. The training rate field, training RMS exit criteria field, and the number of training iterations were set to 0.2, 0.1, and 100, respectively.

4) Support Vector Machine (SVM)

To classify the burned area, a Support Vector Machine (SVM) has been applied to KOMPSAT-3A and Sentinel-2 imagery. SVM is a machine learning algorithm and is a supervised classification method that requires predefined training data (Cortes and Vapnik, 1995). SVM approach seeks to find the optimal hyperplane that separates classes by focusing on the training cases that lie at the edge of the class distributions, the support vectors, with the other training cases effectively discarded (Kadavi and Lee, 2018; Mercier and Lennon, 2003; Zulpe and Pawar, 2012). Thus, an optimal hyperplane fitted, and the approach may be expected to get high accuracy with few training sets, which could be a very advantageous feature given the costs of training data acquisition sets. Therefore, the basis of the SVM approach to classification is the notion that only the training samples that lie on the class boundaries are necessary for discrimination (Foody and Mathur, 2004).

Many hyperplanes could be fitted to separate the classes, but there is only one optimal separating hyperplane, which is expected to generalize well compared to other hyperplanes. This optimal hyperplane should run between two classes, with all cases of a class located to one side of the separating hyperplane, which is itself located at the closest training data points in both classes is as large as possible. Fig. 4 shows the basics classification with SVM; the training data points on these two hyperplanes are called support vectors and are central to establish the optimal hyperplane. The support vectors of the two classes line on two hyperplanes, parallel to the optimal hyperplane (in the middle). The optimal hyperplane separates two classes (red and blue diamond) and is located at the largest margin.

Fig. 4. Schematic depicting the outcome of the SVM algorithm (modified from Foody and Mathur, 2004).

When implementing an SVM, it is vital to select an appropriate kernel (Tehrany et al., 2014). SVM classifiers feature four types of common kernels: linear, polynomial, radial basis function (RBF), and sigmoid kernels (Liu et al., 2013). The RBF kernel works well in most instances, facilitating excellent non-linear classification (Mountrakis et al., 2011; Zhai and Jiang, 2014); therefore, we used this kernel. The gamma value was set to 0.05. The penalty parameter (which controls the tradeoff between acceptable training errors and the forcing of rigid margins) was set to 100 to create the most accurate possible model. The pyramid parameter was set to zero so that each image would be processed at full resolution. Finally, we chose a classification probability threshold of zero to force a class label assignment to all image pixels; we did not wish to evaluate unclassified pixels.

5) Accuracy Assessment

The error matrix was utilized to assess accuracy, which guaranteed the quality of information derived from the remotely sensed data (Kim, 2016). The error is carried out by comparing the results from image classification to ground truth data; these data are typically represented by sample points (Congalton, 1991; Foody, 2002). Initially, 25 random training samples (polygons) homogeneous in terms of land cover type were selected (five for each class), and 297 points of data were selected as test data. Training and test data for four classes (urban, farm, water, and forest) were generated from a land cover map published by the Korea Institute of Geoscience and Mineral Resources (KIGAM) and for burned area class was generated from Normalized Burned Ratio (NBR) map generated from Sentinel-2. The error matrix displays a detailed assessment of the agreement between the reference data and classified results, which indicates the occurrence of misclassification (Kim, 2016). To demonstrate the accuracy evaluation, the overall accuracy and kappa coefficient calculated from the error matrix are used (Cohen, 1960). The overall accuracy is determined by the total number of test data or sample points, and the kappa coefficient is an overall measurement of the error matrix’s statistical agreement. The kappa is a measure of the proportionate reduction in error, i.e., the extent to which the results indicate statistical non-independence. The kappa adjusts for some of the differences among matrices and can compare results for different regions or those of different classifications (Khorram, 1999). Thus, we calculated the kappa statistic of each error matrix to assess if one classifier was significantly better than the other.

4. Result

1) KOMPSAT-3A Image Classification

Post-wildfire KOMPSAT-3A image results were successfully produced by the ANN, as shown in Fig. 5(a), while by the SVM shown in Fig. 5(b). Stratified random sampling on a pixel-by-pixel basis was used for classification, which identified five classes for post- wildfire. There are burned areas (red), water (blue), urban (cyan), farm (yellow), and forest (green).

 

 

Fig. 5. Results of (a) the ANN and (b) the SVM classification from KOMPSAT-3A of Sokcho wildfire, The data were acquired on April 20, 2019.

The ANN and SVM classifiers yielded similar post- wildfire map results. The post-wildfire map results between the ANN and SVM from the KOMPSAT-3A image were well classified. Both ANN and SVM can classify the five classes assigned from the training data. However, the spectral similarity from the satellite image between three classes: urban, farm, and burned area, resulted in interchangeably-misclassification in some parts of both ANN and SVM. Some areas inside the burned area were classified as farms; likewise, some farm areas were classified as burned areas. These can be notified in the north part of the wildfire zone in Goseong County. As seen in Fig. 2(a), the spectral between paddy field area and burned area was similar, so the paddy field areas were detected as burned areas (dark green) by ANN classifier. The SVM classification result did not misclassify the northern area. Unfortunately, many pixels inside burned areas were classified as farms. These results show that different approaches toward the spectral similarity between classes result in different outcomes of classification. Therefore, an accuracy assessment is needed.

The post-wildfire maps produced by ANN and SVM appear similar in many ways, but there are some notable differences in particular areas. For example, the burned area in the ANN classification was broader, as demonstrated by the number of pixels (1, 935, 587), than the burned area in the SVM classification (1, 716, 877). While the ANN classification also showed a larger area of urban area, which differed by 438.289 pixels from the SVM classification. However, the ANN classified a smaller farm area (2, 226, 569 pixels) than the SVM classification (2, 913, 474 pixels). Nevertheless, both results were flawed due to misclassification of some areas; an accuracy assessment was therefore performed to address this issue and is described in the following section. Accuracy assessment also reveals which model outperformed in mapping post-wildfire burned areas.

2) Accuracy Assessment for KOMPSAT-3A Image Classification

Table 2 and 3 show the error matrices of post-wildfire maps in Sokcho derived using the two methods. Test data (297 points) were used to generate the accuracies of SVM and ANN. Of the two classifiers, the SVM was slightly more accurate than ANN, as shown by the higher overall accuracy and kappa coefficient. The SVM algorithm’s overall accuracy and kappa coefficients were 95.29 and 0.940, whereas those from the ANN algorithm were 94.61% and 0.931, respectively. Thus, the SVM yielded higher accuracy than ANN for KOMPSAT-3A image classification.

 Table 2. Error matrix on classified map of burned area from KOMPSAT-3A using ANN method

Table 3. Error matrix on classified map of burned area from KOMPSAT-3A using SVM method

3) Sentinel-2 Image Classification

In this study, post-wildfire Sentinel-2 imagery was also used to map burned area from Sokcho wildfire. Similar to KOMPSAT-3A classification, the Sentinel- 2 image was classified using the ANN and SVM models. Same training data used for KOMPSAT-3A were also used for training in Sentinel-2 classification. Post-wildfire classification maps were successfully produced by the ANN, as shown in Fig. 6(a), while by the SVM shown in Fig. 6(b). There are burned area (red), water (blue), urban (cyan), farm (yellow), and forest (green).

Fig. 6. Results of (a) the ANN and (b) the SVM classification from Sentinel-2 of Sokcho wildfire. The data were acquired on April 20, 2019.

Like KOMPSAT-3A classification results, the ANN and the SVM classifier generated similar post-wildfire maps from Sentinel-2 results. Also, the spectral’s similarity from the satellite image between three classes: burned area, farm, and urban. However, both ANN and SVM classified some farm areas as burned areas in the north wildfire zone near Goseong County. The southern parts of both maps show that some urban areas were classified as farms. From the two maps, the differences between the ANN and SVM for Sentinel-2 were also assessed by the error matrices accuracy assessment in the following section.

4) Accuracy Assessment for Sentinel-2 Image Classification

After the post-wildfire classification had been acquired from the Sentinel-2 imagery, an accuracy assessment was performed using the same data and method used for KOMPSAT-3A accuracy assessment. 297 points were used as the test data to generate the ANN and SVM accuracy assessment. The accuracy results for the ANN and SVM are shown in Table 4 and 5, respectively.

 Table 4. Error matrix on classified map of burned area from Sentinel-2 using ANN method

Table 5. Error matrix on classified map of burned area from Sentinel-2 using SVM method

According to the two algorithms’ accuracy assessment, the overall accuracy and kappa coefficient indicated generally accurate results. The post-wildfire overall accuracy and kappa coefficient from the ANN algorithm reached 90.11% and 0.875, while those for the SVM were 91.52% and 0.893, respectively. The overall accuracy and kappa coefficient from the SVM algorithm were slightly higher than the ANN.

5. Discussion

There was a need to classify the burned area from the Sokcho wildfire in 2019. By collecting data from KOMPSAT-3A and Sentinel-2, post-wildfire maps were generated. Two classification methods were used to produce the post-wildfire maps: the ANN algorithm and the SVM algorithm, which is considered a novel method for mapping burned area, especially in the region of Sokcho. To reveal the accuracy of the two methods, an accuracy assessment was performed. The two satellite imageries were trained and tested, and satisfactory results were obtained.

In general, the ANN and SVM models successfully distinguished the classes in post-wildfire events with high accuracy. By comparing the results from both satellite imageries, the SVM was slightly better than the SVM algorithm. For the post-wildfire by KOMPSAT-3A results, the ANN and SVM differed by 0.68% and 0.009 in overall accuracy and kappa coefficient, respectively; for the Sentinel-2 results, the ANN and SVM differed by 1.41% in overall accuracy and 0.017 in kappa coefficient. The accuracy of the detected fire area from ANN and SVM are not significantly different. However, for processing Sentinel-2 images, SVM only spent about 10 minutes, while ANN spent more than 1 hour. The ANN consumes more time than the SVM because ANN iterates the training data about 100 times to find the hidden layer. On the other hand, SVM only a process to find the best hyperplane to distinguish two classes. Therefore, with less time-consuming processing and better accuracy, SVM is considered a better method to map burned areas. KOMPSAT-3A (2.2 m GSD) has better accuracy in both ANN and SVM results than Sentinel-2 (10 m GSD). These indicate that spatial resolution substantially affects the classification. Thus, high-resolution images afford better image classification in agreement with other studies (Kadavi and Lee, 2018).

The total burned area in the Sokcho wildfire was compared with the result from Korea Forest Service, which calculated through field investigation was 1227 ha or 1, 227, 000 m²(ESRI, 2019); this was similar to KOMPSAT-3A results. The total burned area from both satellite data detected by ANN and SVM ranged from 10, 898, 881 m²to 12, 287, 271 m²for KOMPSAT-3A and 10, 726, 122 m²to 11, 590, 783 m²for Sentinel-2. The difference in the total burned areas might be due to various factors, e.g., misclassification, different spatial resolution. KOMPSAT-3A image classification acquired on April 20, 2019, showed a similar result with field investigation by Korea Forest Service on May 1, 2019. By using the KOMPSAT-3A satellite, the burned area map could be generated faster and has similar results with field investigation. Therefore, the fire mitigation, management, and evaluation could be done earlier, also prevent further damage and restore the area affected by the wildfire (soil leakage and landslide). KOMPSAT- 3A regularly revisits within 1.4 days has the advantage to provide more recent data after a disaster happened. Also, KOMPSAT-3A operates by KARI; therefore, the Korean government could easily access and use this satellite for mapping disaster in Korea.

Additionally, by considering ANN and SVM’s accuracy, several aspects of the differences in KOMPSAT-3A and Sentinel-2 results can be explored. For example, in the classifying process, some areas had a similar spectral that confused the classifier when generating the output class, such as some agriculture areas similar to burned areas or urban areas recognized as farm areas. This problem presented a challenge for the classification of the optical image using a pixel-by- pixel based approach. The availability and cloud cover factor of the satellite imagery used were likely to have affected the acquisition of soil or vegetation data by satellite. The late date of acquisition, relative to the event, indicates that conditions may differ from those initially after the wildfire. Because of these limitations, the collaboration of materials is needed for future studies. Nevertheless, this study is relevant, and the findings can be applied to other events in other regions.

6. Conclusion

This study demonstrates that the methods used, ANN and SVM could reveal the burned area map for the Sokcho wildfire in 2019. The SVM showed higher accuracy (overall accuracy 95.29%) compared with ANN (overall accuracy of 94.61%) for the KOMPSAT- 3A. Moreover, for Sentinel-2, the SVM attained a higher accuracy (overall accuracy of 91.52%) than the ANN algorithm (overall accuracy 90.11%). Notably, the satellite data and classifier used influenced the results of classification and accuracy. It is crucial to map the burned area in disaster assessment after a wildfire to prevent further damage and restore the affected area. Thus, this study’s results are expected to be useful for researchers, planners, and policymakers as a basis for analyzing the occurrence of wildfire in Korea and developing mitigation and evacuation plans to minimize the severity of such hazards in the future. Although the results managed to map the burned area to mitigate an evacuation, a different approach might be preferable in future studies such as optimization or deep learning algorithm to achieve the most effective mapping of the burned area, especially in Korea.