As global climate change intensifies precipitation variability, the risk of flood damage is increasing. Rising temperatures boost the atmospheric water vapor content, alter hydrological patterns, and increase the frequency of flooding events (Tabari, 2020). The Korean Peninsula, for example, endured its longest recorded rainy season in the summer of 2020, leading to the designation of 38 cities and counties as special disaster zones and 14 deaths (Lee et al., 2020). According to the 6th Intergovernmental Panel on Climate Change (IPCC) report, the National Institute of Meteorological Sciences projected a precipitation increase of over 3% and a rise in maximum precipitation exceeding 14% over five days in East Asia (Kwon et al., 2020). Additionally, a 2022 press release from the Climate Change Monitoring Division (2022) predicted that even under low-carbon scenarios, the frequency of extreme precipitation events that occur once a century in Korea will increase by 31% in the first half of the 21st century.

The anticipated damage from climate change extends to North Korea as well. According to Kang et al. (2021), the frequency of floods in North Korea has been on the rise, with eight floods reported in the 1990s, 13 in the 2000s, and a significant increase to 20 in the 2010s. Kim (2011) noted significant damage to agricultural infrastructure in major grain-producing regions, such as Hwanghae-do and Pyeongan-do, primarily due to heavy rains during the critical rice growth period. In 2011, Typhoon Echo resulted in the flooding of over 14,876 hectares of agricultural land and damaged at least 3,000 houses by July 15. The International Federation of Red Cross and Red Crescent Societies (IFRC) reported that in 2020 alone, more than 8,256 houses were destroyed and over 22,000 hectares of agricultural land were devastated across North Korea due to floods (Ahn, 2020). Furthermore, the Economic and Social Commission for Asia and the Pacific (ESCAP, 2023) reported in 2023 that between 1991 and 2020, floods and storms claimed the lives of 2,698 individuals in North Korea. Despite the implementation of the National Land Management Mobilization Project in 1996, which was aimed at disaster prevention, numerous issues persist due to insufficient support, Kim (2022) highlighting the limited capacity of North Korean authorities to manage natural disasters, including floods.

Analyzing flood damage in North Korea presents substantial challenges because of significant access restrictions in the country. Although North Korea shares observational data at 27 points through the World Meteorological Organization’s (WMO) weather communication network, its self-reported precipitation data is incomplete in terms of time-series coverage. Additionally, its weather observation technology is assessed at a level comparable to that of South Korea in the 1990s, diminishing the overall reliability of the data (Kim et al., 2020). Moreover, the 2022 INFORM Risk Index report further highlighted the limited analytical capacity of North Korea (Thow et al., 2022), as evidenced by its governance score of 8.0 points and communication system infrastructure score of 7.3 points. These scores reflect a significant deficiency in the ability of the country to respond effectively to natural disasters. Moreover, financial constraints, an ineffective legal system, organizational inefficiencies, a shortage of disaster experts, and the absence of a comprehensive disaster database exacerbate the challenges faced by North Korea in mitigating flood damage (Hong, 2020).

In regions with restricted access, remote sensing through satellite technology, particularly synthetic aperture radar (SAR), is essential for analyzing flood damage. SAR has the distinct advantage of being usable under all-weather conditions, as it operates effectively regardless of cloud cover by using longwavelength electromagnetic waves. Lim and Lee (2018) employed Sentinel-1 SAR data to measure flooded areas in North Korea by comparing images taken before and after the August 2016 flood. Moreover, in another study (Kim et al., 2019), the flooded area in Yeongdeok-gun affected by Typhoon Kong-rey in 2018 was identified using optical images of Sentinel-1 ground range detected (GRD) and PlanetScope, with the analysis supported by CCTV data from the location as validation data. A common technique for SAR image-based analysis is image differencing, which involves detecting changes through the subtraction of two images. This method visually emphasizes differences based on radiation values, resulting in a value of zero for identical areas and negative or positive values indicating changes. The combination of bands emphasizes visual differences based on radiation values (Goswami et al., 2022). However, even studies from the 2020s utilizing machine learning have largely focused on water body detection rather than analyzing changes between different periods (Choi et al., 2022; Lee et al., 2022). Consequently, it is difficult to define difference images as a technique actively used for flood detection.

Therefore, this study detected flood damage in North Korea in 2023 by using the image differencing technique on Sentinel-1 SAR GRD images. The results demonstrate the viability of determining flooded areas by applying artificial intelligence (AI) classification methods to the image differencing technique. Additionally, this approach facilitates the assessment of the vulnerability level to floods in North Korea. It is anticipated that these findings will enhance research methodologies for analyzing flood damage and inundation areas in regions with limited access, including North Korea, in the future.

2.1. Study Area

In this study, the focus was on the interiors of five rectangular areas within North Korea that were identified as flood-affected in 2023 based on media reports and weather data (Fig. 1). Given the characteristics of North Korea, where geographic information, including watershed maps, is not sufficiently accessible, and considering the methodology of this study aimed at estimating the socio-cultural impacts of specific events, it was deemed more appropriate to directly designate the area of interest rather than conducting investigations based on watershed units. Also, considering the lack of publicly available data on population distribution and infrastructure in North Korea, using specific administrative boundaries is unlikely to be useful for meaningful research. Therefore, it was deemed more appropriate to define target areas by directly delineating polygons based on regions where specific events could be confirmed through reports. According to the Korea Meteorological Administration, North Korea typically receives an average annual precipitation of 912.0 mm, with 543.2 mm occurring in summer months (Korea Meteorological Administration, 2024b). Although this is less than the over 1,100.0 mm average annual precipitation in the central regions of South Korea (Korea Meteorological Administration, 2024a), the five target sites are situated in the southern part of North Korea.

Figure 1. Location map and SAR image of study areas.

Consequently, it is speculated that the rainfall patterns in these areas resemble those observed in South Korea. The study targeted areas affected by heavy rainfall events. Areas 1 and 4, as reported by Seoul Pyongyang News (Ahn, 2023a), experienced significant rainfall, with Pangyo and Taetan recording over 200 mm from August 21 to 23, 2023. Area 2 was selected based on a report from Radio Free Asia (Han and Cheon, 2023), which highlighted that Sentinel-2B imagery captured a cooperative farm in Yongmadong, Cheongdan-gun, Hwanghaenam-do, which was flooded by heavy rain on July 19, 2023. Area 3 includes the path of Typhoon Kanun, which affected areas such as Geumcheon-gun and Singyeeup before dissipating in 2023. Lastly, Area 5 was selected to cover regions like Pangyo-gun, Cell-gun, and Pyeonggang-gun, following the report of Seoul Pyongyang News of heavy rainfall in Gangwondo and Hwanghae-do on July 4 and 5, 2023 (Ahn, 2023b). A detailed summary of the coordinates and widths of each target site is provided in Table 1.

Table 1 . Spatial characteristics and the shooting dates of study areas.

Area	Latitude	Longitude	Area (km2)	Scene 1 date	Scene 2 date
Area 1	38.044009	125.291104	33.49	2023.08.16	2023.08.28
	38.101713	125.337796
Area 2	37.872865	125.894732	76.19	2023.06.29	2023.07.23
	37.929484	126.002879
Area 3	38.169551	126.405467	1,933.69	2023.08.04	2023.08.16
	38.533728	126.834193
Area 4	38.698536	126.920664	204.96	2023.08.16	2023.08.28
	38.813125	127.069667
Area 5	38.701977	127.124261	211.82	2023.07.04	2023.07.16
	38.779099	127.338494

2.2. Dataset

This study utilized SAR GRD images from the Sentinel-1 satellite, provided by the European Space Agency (ESA). Sentinel-1 is particularly effective for analyzing the target site due to its spatial resolution of 10 m and temporal resolution of 6 days. The dataset period starts from October 3, 2014. The provided bands consist of four combinations of horizontal and vertical polarization: HH, HV, VV, and VH. This study primarily employed VV band images as the primary research data, following insights from previous studies (Jeong et al., 2021), which highlighted the ability of VV polarized data to distinguish between aqueous and nonaqueous regions because of its sensitivity to surface roughness. Since the intervals between scenes are all less than four weeks and fall within the same season, it is considered that the confusion between shadows and water bodies is unlikely to have a significant impact on the research results. This study constructed different images using the Google Earth Engine. The VV band images went through the preprocessing steps of thermal noise removal, radial calibration, and terrain correction. The image acquisition dates for two images used in the image differencing construction are detailed in Table 1.

To analyze the flood characteristics and impacts at the target site, understanding the land use within the flooded area is crucial. For this purpose, this study used the Land Use/Land Cover (LULC) map of December 31, 2022, from the latest revision of the 10 m LULC Time Series provided by the Sentinel-2 satellite, sourced from ESRI. This cover map is particularly advantageous for this study as it matches the 10 m spatial resolution of the Sentinel-1 SAR GRD image. The LULC map offers a detailed classification into nine classes: water, trees, flood vegetation, crops, built area, bare ground, snow/ice, clouds, and rangeland (Fig. 2). This categorization is crucial for this study because it enables the extraction of agricultural and developed lands, providing a thorough understanding of the land cover within the areas impacted by flooding.

Figure 2. Land Use/Land Cover (LULC) map of the study area: (a) Area 1, (b) Area 2, (c) Area 3, (d) Area 4, and (e) Area 5 (source: Esri, Sentinel-2 Land Cover Explorer).

2.3. Method

This study used random forest classification based on the visual differences highlighted by the reflection values in the image differencing results. In this study, the random forest classification technique, a widely used AI-supervised classification method known for its high accuracy and robust performance, was employed (Breiman, 2021). Kang and Lim (2023) previously analyzed land cover changes using Random Forest in their study on coastal changes in North Korea, providing a precedent that supports its suitability for use in this study. The Random Trees function in ArcGIS Pro was utilized for this study. The maximum number of trees and depth were set to the default values of 50 and 30, respectively. Training data for target site extraction were selected from images of Unmun Dam, Hoengseong Dam, and Andong Dam in South Korea, where publicly available hydrological information helped identify changes in water quantity and distinguish between water and land.

The choice of image period for creating the different images was strategically made to maximize the observable difference in water level between the two images, based on water level data from the water environment information system under the Ministry of Environment. The details regarding the location and timing of the image used for data extraction at the target sites are detailed in Fig. 4. In the image differencing result, compared to Scene 1, pixels representing newly formed water are indicated in red, areas where water has receded are indicated in blue, unchanged water pixels are represented in black, and unchanged land pixels are indicated in white (as shown in Fig. 3). Based on these color differences, the analysis target sites in North Korea and the data extraction target sites in South Korea were classified into four classes: water, flooding, land, and drainage. Subsequently, the Clip tool in ArcGIS Pro was used to calculate the areas of the ESRI land cover map, flooded cropland, and developed land in the target analysis site. It is important to note that in the optical images, both crops and rangeland were classified under cropland due to the minimal significance of their differences.

Figure 3. Sentinel-1 SAR GRD image of Area 5.

Figure 4. Location and shooting dates of training sampling sites.

The accuracy assessment tool in ArcGIS Pro was employed to validate the classification results. Validation data were derived from images identical to the training data. For extracting training data, references were made to satellite images, including the target site, obtained during the research period using Google Earth Pro, which offers optical images captured by high-resolution satellites and is widely used as auxiliary data in remote sensing applications, particularly when dealing with geographic information in North Korea, where access is restricted (Kim, 2022). Previous studies, such as (Yoon et al., 2018), examining mining activities, and (Ki, 2016) analyzing deforestation in North Korea, have successfully employed Google Earth images for detailed analyses. Based on these precedents, Google Earth images were deemed appropriate for use as auxiliary data in this study, especially for distinguishing between water and land pixels. In addition to Google Earth images, high-resolution images from Maxar Technologies and CNES Airbus, which covered the target sites, were referenced. Subsequently, an accuracy evaluation was conducted on the random trees classification image of the data extraction target site, and a confusion matrix was generated to quantitatively assess the performance of the classification.

3.1. Detecting Flooded Area

This study used random forest classification to classify the data into four classes, as shown in Fig. 5 and analyzed the area of each classification group. This analysis enabled the calculation of the flooded land area, which is a central focus of this study. The flooded land areas at the target sites and the flooding ratios for croplands and developed lands are detailed in Table 2.

Figure 5. Four classes generated by random trees classification: (a) Area 1, (b) Area 2, (c) Area 3, (d) Area 4, and (e) Area 5.

Table 2 . Areas of classes and flooded regions.

Area	Area 1	Area 2	Area 3	Area 4	Area 5
Cropland (km2)	32.8637	91.8417	1,805.077	172.2809	120.9544
Developed (km2)	3.8304	0.3198	15.5429	2.3643	0.3389
Flooded cropland area	0.1283	1.6139	7.8117	1.0643	2.1072
Flooded developed area (km2)	0.0344	0.0020	0.0521	0.0215	0.0010
Ratio of flooded area in cropland (%)	0.3904	1.7573	0.4328	0.6178	1.7421
Ratio of flooded areas in developed (%)	0.8981	0.6254	0.3352	0.9094	0.2951

From August 16 to 28, 2023, the study recorded 1.1926 km² of cropland and 0.0559 km² of developed land flooded in the target areas. Typhoon Kanun caused flooding of 7.8117 km² of cropland and 0.0521 km² of developed land in Area 3. In Area 2, which includes a cooperative farm, the analysis revealed that 1.6139 km² of cropland and 0.0020 km² of developed land were flooded. Area 5 saw 2.1072 km² of cropland and 0.0010 km² of developed land flooded. The flooding ratios of developed land in Areas 1 and 4 were higher than that of cropland. In Areas 5 and 2, the latter which includes a cooperative farm, the flooding rate of cropland was more than twice that of developed land. In Area 3, the most extensively damaged region, the flooding rate of cropland was over 29% higher than that of the developed land.

3.2. Accuracy Assessment

As discussed in Section 2.3, accuracy validation involved constructing a confusion matrix and calculating the kappa coefficient. The kappa coefficient is a statistical tool used to assess the performance of classification models, and it is especially useful for multiclass classification problems. The kappa coefficient measures the performance of the classification methodology by comparing the observed agreement with the agreement expected by chance. The resulting confusion matrix is shown in Table 3. User Accuracy (Type 1 error, U_Accuracy) recorded a perfect score of 1 for the drainage class, indicating no errors, while the lowest accuracy for flooding was still high at 0.9167, exceeding 0.9 overall. Producer Accuracy (Type 2 error, P_Accuracy) reached its highest value at 0.9918 for land, but the lowest accuracy for drainage was calculated as 0.5. The calculated kappa value of 0.9506 reflects high overall accuracy in the classification.

Table 3 . Confusion matrix.

Class value	C_1	C_2	C_3	C_4	Total	U_Accuracy	Kappa
C_1 - Water	97	1	1	1	100	0.97	0
C_2 - Flood	1	33	2	0	36	0.9167	0
C_3 - Land	0	1	363	4	368	0.9864	0
C_4 - Drainage	0	0	0	5	5	1	0
Total	98	35	366	10	509	0	0
P_Accuracy	0.9898	0.9429	0.9918	0.5	0	0.9784	0
Kappa	0	0	0	0	0	0	0.9506

Except for Area 1, all other target areas were found to have experienced flood damage exceeding 1 km². Compared to previous flood cases in specific North Korean regions, the impact of Typhoon Kanun in Area 3 was significant, with 7.8117 km² of cropland flooded. This exceeds the damage caused by Typhoon Maisak in 2016, which flooded 1.28 km² of cropland in Hoeryeong-si and Eunseong-gun, Hamgyeongbuk-do (Kim et al., 2019a), and Typhoon Goni in 2015, which destroyed 1.24 km² of cropland in Nason-si (Ministry of Unification, 2024), resulting in over 40 fatalities. This flooding from Typhoon Kanun stands out among flood cases caused by typhoons in a single area over the past decade. In contrast, according to the Ministry of Unification, when compared with flooding from just heavy rainfall, the inundation damage at the target sites was less severe than the catastrophic flood in Kaesong in 2008, where 15.20 km² of cropland was flooded. This pattern was also evident in the flooding in South Hamgyeong Province in August 2021, which destroyed 1,170 homes. The relatively lower extent of flood damage in Hamgyeongnam-do in 2021 and in Areas 1, 2, 4, and 5 compared to the Kaesong event in 2008 points to improved flood response capabilities in North Korea, indicating advancements since the 2000s.

Compared with South Korea during the same period, the flooding from Typhoon Kanun in 2023 in North Korea 2023, which flooded 6.53 km² of cropland in Gyeongsangbuk-do and 1.58 km² in Jeju (Lee, 2023), covered a larger area of flooded land. This comparison emphasizes the greater vulnerability of North Korea to typhoon-induced flood damage than South Korea. Conversely, when compared with South Korea where 5.32 km² were damaged in Mungyeong City and 7.60 km² in Cheongyanggun in July 2023, the damaged area was smaller in Areas 2 and 5, which experienced flooding during a similar period. This indicates the need for further research to explore and compare the vulnerability of flood damage from simple heavy rains between the two countries.

The period from July to August, which coincided with the flooding at the target site, was also a critical period for the safe harvesting of rice in North Korea (Yang et al., 2018). In particular, in Hwanghae-do, where Area 2 is located, many agricultural facilities are vulnerable to damage from adverse weather conditions (Kim et al., 2023). Given this, the flood damage to cropland analyzed in this study is seen as significant and likely to present a major challenge for North Korean authorities. Additionally, it is vital to consider the potential consequences of Typhoon Kanun, which flooded a larger agricultural area than Typhoon Goni did in Rason City, where it caused over 40 deaths. Furthermore, comparing the flooded area in this study with other flood cases in North Korea allows for an estimation of the property damage incurred during the study period based on official announcements and market prices in North Korea. It is important to note that comparing the vulnerability of cropland and developed land may be challenging based solely on the ratio of flooded areas.

The accuracy evaluation demonstrated that both U_Accuracy and P_Accuracy for water, flooding, and land were 0.9 or higher, confirming the suitability of the methodology used in this study for detecting flooding. However, for drainage, while U_Accuracy reached 1, indicating no type 1 errors, P_Accuracy was only 0.5. flooding recorded the lowest U_Accuracy and the second lowest P_Accuracy after drainage. This result likely stems from the misidentification of artificial changes, such as the installation of water supply infrastructure, irrigation, and artificial structures, as flooding and drainage caused by weather factors. The calculated kappa value of 0.9506 further demonstrates the high accuracy of the methodology used in this study. When compared with previous studies on SAR images, the methodology in this study is considered appropriate, as evidenced by a kappa value comparable to the results of VV polarization image classification in (Cao et al., 2019), which reported a kappa value of up to 0.91.

Based on the information provided, it is concluded that using SAR imagery and change detection techniques is effective for identifying the affected areas and assessing response levels and effectiveness. This approach overcomes the limitations posed by cloud cover and adverse weather conditions common during flood-prone seasons, such as those caused by the monsoon climate. Sentinel-1 SAR-based flood monitoring is well suited for remote sensing in regions with limited accessibility, such as North Korea and other developing countries. Additionally, this method can serve as a reference for managing similar conditions of flood damage and weather-related disasters.

In this study, the flooded area of the target site in North Korea, which is believed to have been flooded in 2023, was calculated at five specific locations using Sentinel-1 SAR GRD images. The level of flooding damage at the target site was considered significant compared to previous flood incidents in North Korea. Additionally, the reduced extent of flood damage from heavy rainfall compared to the 2000s indicates that North Korea has improved its capacity to respond to such events. Compared with South Korea during the same period, Typhoon Kanun caused more substantial damage. The accuracy assessment yielded a high kappa value of 0.9506, indicating that this study successfully conducted flood analysis for the target area. By comparing the flooded area with previous flood cases in North Korea, it is possible to estimate casualties and property damage at the target site. This demonstrates that SAR-based change detection by remote sensing is suitable for flood monitoring in regions with limited accessibility. The results also suggest that these techniques can serve as reference data when responding to flood damage.

However, the observed lower accuracy for drainage and flooding compared to water and land suggests a challenge in distinguishing artificial changes, such as irrigation, draining, and civil engineering works, from flood damage. In regions with extremely limited socio-cultural and geographical data, such as North Korea, securing reliable label data and defining appropriate study areas remains a persistent challenge. Future research should employ high resolution satellites directly and integrate multiple satellite data to improve the spatial and temporal resolution of image datasets. This approach will enable the generation of more accurate training and Validation data, and the use of reliable land cover maps. Enhancing data quality and resolution will ultimately improve the accuracy and reliability of flood damage detection and analysis.

This research was funded by a National Research Foundation of Korea grant provided by the Ministry of Science and ICT (No. 2022R1C1C1008489) and a Kookmin University grant.

No potential conflict of interest relevant to this article was reported.