Global warming has emerged as a critical global issue, primarily driven by carbon dioxide emissions and climate change (Intergovernmental Panel on Climate Change, 2007). Among the key contributors to this accelerating trend, the rise in carbon dioxide emissions constitutes a substantial portion. Forests, as vital carbon sinks in the context of climate change, play an essential role in absorbing atmospheric carbon (Feng et al., 2022). Consequently, the need for accurate forest resource information to enable effective management is increasingly critical, and the ability to respond promptly and efficiently to land use changes in forested areas is becoming more essential (Ameray et al., 2021).

There are several factors that affect the change in forest land use (Negassa et al., 2020); large-scale forest damage is increasingly attributed to abnormal weather events, including hail and drought. Forest fires, in particular, can drastically alter large areas over a short duration (Tyukavina et al., 2022). In the event of such damage, the government and relevant organizations typically undertake joint field investigations to assess the impacted areas. However, in cases where forests are predominantly situated in mountainous regions, limited accessibility poses significant challenges to rapid response efforts; field investigation methods are inherently limited by accessibility constraints. While unmanned aerial vehicles (UAVs) have been introduced to survey larger areas (Diez et al., 2021), they remain constrained in their ability to analyze large-scale land use changes over time. In this context, satellite imagery offers a highly effective solution, allowing for periodic evaluations of large-scale areas, particularly with significant temporal and spatial variability (Decuyper et al., 2022).

Consequently, ongoing research seeks to evaluate the current status of forests by utilizing a range of domestic and international satellite data (Abbas et al., 2020; Waśniewski et al., 2020). In particular, studies on forest vegetation have actively utilized vegetation indices from satellite imagery to detect and monitor forests over large areas (Lemenkova and Debeir, 2023). Typically, the normalized difference vegetation index (NDVI) has been the primary tool, as it is the most widely used vegetation index for identifying the biological characteristics of vegetation (Jiang et al., 2006). In order to assess the changes in forest vegetation caused by wildfires, numerous studies have employed vegetation-related indices, given that vegetation activity typically declines following such events (Ba et al., 2022). For example, desertification in the Ordos region of China was analyzed using NDVI (Xu et al., 2009), while NDVI was used to examine vegetation trends in Mongolia (Meng et al., 2020). Additionally, modified NDVI has been utilized to detect deforested areas in North Korea based on satellite imagery (Lee et al., 2016). It is well known that rapid and cost-effective forest monitoring across wide areas can be conducted using vegetation indices based on multispectral imagery such as Landsat and Sentinel-2, which provides global coverage (Jönsson et al., 2018).

However, in the case of optical satellite imagery, spectral characteristics may be affected by humidity, weather conditions, or cloud cover (Richter, 1996). Consequently, NDVI values derived from imagery captured at different time intervals may not consistently reflect the same level of vegetation vitality. This variability suggests that the threshold for determining forest presence may differ between images taken at different periods. Generally, it is understood that tropical and temperate forests exhibit NDVI values of 0.6 or higher (Cai et al., 2014), rather than having a specific value as a standard. Given this variation, it is challenging to rely on a single NDVI value to accurately assess forest presence across images from different periods. Therefore, a method that effectively leverages the relative differences in NDVI values within single-period images is needed to enhance the accuracy of forest detection.

In particular, the CAS500-4, which is known as the agriculture and forestry satellite and is set to begin operations in 2025, will significantly enhance satellite monitoring capabilities in Korea. In addition to the standard R, G, and B bands, the satellite is equipped with Red Edge (RE) and Near Infrared (NIR) sensors, which are highly sensitive to vegetation vitality (Kwon et al., 2021). With its ability to capture imagery of the entire country on a three-day cycle, the CAS500-4 is expected to provide optimized data for monitoring forest changes on a national scale.

Therefore, this study proposes a method for the automated detection of primary forest changes across large areas by utilizing multi-temporal, mid-resolution satellite imagery. Specifically, in order to pre-establish training data for forest areas, forest area information of digital forest map was used which is generated from high-resolution data sources such as field surveys and aerial photographs; and a probability map of forests for each period is to be created based on machine learning in order to quantitatively detect forest changes. This approach will normalize satellite images from various periods to facilitate direct comparison and allow for the identification of changes in forest probability values over time. By analyzing multi-temporal satellite data, the method focuses on detecting land use changes in existing forested areas. In this study, a forest fire area where rapid forest changes occur was selected as a case study for the proposed methodology. It shows that forest changes across large regions could be automatically monitored by using continuously updated satellite imagery, allowing for the identification of the location and approximate extent of forest damage. The results can serve as a scientific basis for forest restoration efforts and inform future policy decisions.

2.1. Study Area

Korean peninsula can be geographically divided into four regions: the expansive plains in the west, the southwestern mountainous area, the wide basin in the southeast, and the high mountain ranges in the east. It is bordered by the Yellow Sea to the west, the East Sea, the Korea Strait, and the East China Sea to the east and approximately 70% of the country is covered by mountainous regions. In terms of climatic characteristics, it experiences a temperate climate with four distinct seasons, characterized by relatively dry conditions in spring, fall, and winter, and increased humidity during the summer. Due to these climatic patterns, the incidence of forest fires is particularly high during the spring and fall seasons.

This study aims to develop a methodology for the automated detection of forest change across large areas using multi-temporal, medium-resolution satellite imagery. It could be also part of a broader initiative to utilize the CAS500-4 which is known as the Korean agricultural and forestry satellite. To pilot this approach, a case study was conducted in an area significantly affected by forest fires, where rapid forest changes occurred over a short period. Given the challenges of natural restoration in fire-damaged forests, it is crucial to accurately identify the location and extent of the affected areas. By accurately identifying forest damage in remote and hard-to-reach areas, this approach enables quicker and more efficient responses, ultimately enhancing future forest management practices.

The case study area of this study is Dogye-eup, Samcheok-si, located in Gangwon-do, South Korea (Fig. 1). The Gangwon region is characterized by its extensive forested areas and a high frequency of large-scale forest fires (Lee and Lee, 2006), particularly in the spring. In Dogye-eup, a significant forest fire was ignited on May 6, 2017, and was fully extinguished four days later, on May 9. The area consists predominantly of pine forests with fire-prone coniferous trees (Choung et al., 2004). Additionally, it is known for strong winds, particularly valley winds that frequently change direction which complicates the firefighting efforts (Lee and Kim, 2011). The strong winds significantly hindered the accurate deployment of firefighting efforts, resulting in delays in containing the fire. Additionally, the fire broke out on the eastern slopes of Baekdudaegan, a region known for its steep and challenging terrain, which further complicated the suppression efforts. As a result, it took approximately 72 hours to fully extinguish the fire, and more than 765 hectares of forest were reported to have been damaged in the Dogye-eup area of Samcheok-si (Korea Forest Service, 2017)

Figure 1. Map of the case study area.

2.2. Materials

Landsat is a remote sensing satellite system designed for Earth observation, with Landsat 8 operational since its launch in 2013. Since the satellite is a representative long-running series of satellites dedicated to global observation, it offers data well-suited for time-series studies with a 16-day revisit cycle. In order to test the usability of agricultural and forestry satellites in advance, this study was also conducted based on Landsat data in terms of utilizing multi-temporal images with multiple spectral bands.

To detect forest changes resulting from forest fires in the case study area, this study utilized Landsat-8 operational land imager (OLI) imagery of Dogye-eup, Samcheok, from 2016 and 2017. Two sets of images were analyzed: one set captured before the forest fires consisting of spring and summer 2016 images, and the other set taken after the fires from spring and summer 2017 (Table 1). Given the seasonal variability of forests in the region, images from comparable periods were selected to ensure accurate comparisons of forest conditions before and after the fire event (Fig. 2).

Figure 2. RGB images used in this study.

Table 1 . Landsat-8 OLI data of the study.

Date	Sensor	Season	Path	Row
2016/05/12	OLI	Spring	115	34
2016/08/25	OLI	Summer	114	34
2017/05/05	OLI	Spring	114	34
2017/07/27	OLI	Summer	114	34

In addition, this study utilized a 1:5,000 digital forest map, which was developed based on a database of aerial forest photographs and visual analyses conducted by forest experts. The forest areas were initially extracted from the existing forest map to minimize false detections for automatic forest change detection based on satellite imagery. This approach was specifically designed to prevent the misidentification of areas with higher vegetation vitality than forests, such as cornfields or rice paddies due to seasonal variations. Through this, it was able to focus more on forest areas being damaged by identifying changes in areas that were previously forests.

For additional consideration, data from the digital elevation model (DEM) was collected. The DEM data was obtained through NASA’s Shuttle Radar Topography Mission (SRTM) Project, which conducted an 11-day mission from February 11 to 22, 2000. This mission provided elevation data covering the area from 60°N to 56°S latitude, with a spatial resolution of 1 arc second. Two sheets of DEM data were acquired to consider the topographical aspects of the mountainous case study area (Table 2).

Table 2 . SRTM DEM used in this study.

SRTM DEM	Sensor	Resolution	N	E
Left	SRTM	1 arc	37	128
Right	SRTM	1 arc	37	129

The aim of this study is to automate the detection of forest changes using multi-temporal, medium-resolution satellite imagery. For the detailed steps, the preprocessing of Landsat satellite images is performed as the initial step. Subsequently, the R, G, B, NIR, shortwave infrared (SWIR), and thermal infrared (TIR) bands are extracted from the satellite imagery, and both NDVI and normalized difference water index (NDWI) are calculated; slope values are derived from DEM data to construct the necessary input layers. Then, training set of forest and non-forest areas was extracted using digital forest maps developed from field surveys and high-resolution aerial photographs, alongside NDVI values derived from satellite images of the corresponding periods. Next, probability maps for forest cover are then generated for each time period using a maching learning model based on neural networks.

Finally, the degree of forest change is subsequently determined by comparing the forest probability maps created from multi-temporal satellite images, and the reduction in forested areas is estimated through time series analysis. This approach enables the automatic detection of forest changes as additional satellite images from subsequent periods are acquired and compared with those from previous periods. The detailed workflow including the automation concept proposed in this study is illustrated in Fig. 3.

Figure 3. Proposed dataflow of this study.

3.1. Pre-Processing

To effectively utilize time-series vegetation index data from satellite imagery, it is essential to use atmospherically corrected images (Ouaidrari and Vermote, 1999). In this study, atmospheric correction was applied using the correction methodology provided for Landsat 8 OLI images by the United States Geological Survey (USGS). Atmospheric correction was conducted for the R, G, B, NIR, and SWIR bands based on the following equations. After correcting the radiance values, the reflectance values were subsequently calculated (Eqs. 1 and 2).

L_λ: Top of atmosphere (TOA) spectral radiance,

M_L: Band-specific multiplicative rescaling factor from the metadata,

Q_cal: Quantized and calibrated standard product pixel values (DN),

A_L: Band-specific additive rescaling factor from the metadata.

ρλ′: TOA planetary reflectance, without correction for solar angle,

M_p: Band-specific multiplicative rescaling factor from the metadata,

Q_cal: Quantized and calibrated standard product pixel values (DN),

A_p: Band-specific additive rescaling factor from the metadata.

For the TIR band, only band 10 was utilized, and it was converted to radiance following Eq. 1. The radiance values from the TIR band were then converted to brightness temperature (BT) values and used in subsequent analyses (Eq. 3) (Ihlen and Zanter, 2019). The BT values were considered in that surface temperature is closely linked to vegetation vitality and survival (Chuvieco et al., 2004).

T: At-satellite brightness temperature (K),

K₁, K₂: Band-specific thermal conversion constant from the metadata,

L_λ: TOA spectral radiance.

3.2. Generating Input Data

NDVI is a normalized vegetation index that was originally developed by (Rouse et al., 1974) and has since become the most widely utilized index for vegetation monitoring and assessment on a regional to global scale (Huang et al., 2021). The NDVI serves as a crucial technique for determining the presence and vitality of vegetation by quantifying vegetation activity through the unique reflectance characteristics exhibited by plant surfaces (Calera et al., 2001). Vegetation demonstrates specific spectral reflectance properties, most notably absorbing wavelengths within the range of 0.64 to 0.67 μm, while simultaneously reflecting strongly in the NIR region which spans from 0.85 to 0.88 μm (Xu and Guo, 2014). In other words, plants tend to have low reflectance in the red portion of the visible light spectrum while their reflectance in the NIR region is characteristically high.

As a result, the reflectance in vegetated areas is significantly higher in the NIR region compared to the visible region, and NDVI is calculated using this distinct characteristic to differentiate and monitor vegetation across broad spatial scales. NDVI is particularly valuable because it provides a single, easily interpretable value that reveals critical information about both the spatial distribution of vegetation and its biomass (Eq. 4). The formula for calculating NDVI results in a value ranging between -1 and 1, with higher values indicating dense and healthy vegetation. According to the NDVI formula, areas covered by vegetation typically show values close to 1, reflecting vigorous plant growth, whereas non-vegetated areas—such as bodies of water, barren land, ice, and snow—exhibit values around 0 or lower, indicating the absence or minimal presence of vegetation.

ρ_NIR: Reflectance value of NIR band,

ρ_RED: Reflectance value of RED band.

Fig. 4 presents the NDVI images for each period within the case study area. As illustrated in Figs. 4(c, d), which depict the NDVI values for 2017, there is a noticeable reduction in NDVI values in the vicinity of Dogye-ri, located in the upper left portion of the images. This significant decrease in NDVI values suggests a substantial decline in vegetation activity in this region during 2017, likely due to the forest fire event. The lower NDVI values indicate areas where vegetation health and density have been compromised, reflecting the extent of damage caused by the forest fire. These visual representations are critical for understanding the spatial distribution of forest loss and the degree of change over time in the affected areas.

Figure 4. Input layers of NDVI.

The NDWI is a spectral index used to provide information on the moisture content within vegetation, utilizing the difference between reflectance values in the NIR region and the SWIR region, in a manner similar to the NDVI (Gao, 1996). The NDWI was first proposed by Gao in 1996 as a tool for assessing vegetation water content, with the index designed to capture subtle variations in plant moisture levels by comparing the absorption characteristics in the NIR and SWIR wavelengths. This index is particularly valuable for applications related to monitoring drought conditions, assessing plant stress, and analyzing changes in water content across large spatial scales (Gu et al., 2008). The equation used to calculate NDWI is as follows (Eq. 5), and it typically results in values ranging from –1 to 1, where higher values indicate greater water content in the vegetation canopy.

ρ_NIR: Reflectance value of NIR band,

ρ_SWIR: Reflectance value of Red band.

DEM and slope are key factors often utilized in the construction of forest fire risk maps due to their influence on fire behavior and susceptibility (Bui et al., 2017). In this context, the present study incorporated SRTM DEM data for the 37°N 129°E location, resampled to a spatial resolution of 30 meters to match that of the Landsat imagery. Using this resampled DEM, the slope values were calculated and integrated into the analysis; the calculations were performed using ArcGIS 10.3 software. Finally, along with the DEM and slope data, the study also included NDVI and NDWI indices, as well as the R, G, B, NIR, SWIR, and TIR bands from the satellite imagery for each period. These bands, extracted through atmospheric correction in the preprocessing step, were employed as input data for further analysis.

3.3. Modeling with Machine Learning Techniques

As previously discussed, the closer the NDVI value is to 1, the greater the likelihood of vegetation presence. It is generally accepted that sufficient forest cover is indicated when NDVI values exceed 0.6, particularly in tropical and temperate forests (Cai et al., 2014). However, due to the influence of various environmental factors, including atmospheric conditions, the NDVI values derived from images acquired at different time intervals may not consistently reflect the same level of vegetation vitality. This variability complicates the direct comparison and analysis of multi-temporal images, making it difficult to accurately assess vegetation changes over time. To address this issue, the present study employed machine learning algorithms to first calculate the probability of forest for each image period. It enabled a more accurate comparison of multi-temporal images, facilitating the detection and analysis of forest changes despite the temporal variability in NDVI values.

Machine learning techniques are widely utilized modeling methods, capable of addressing complex functions and systems involving numerous variables, such as those found in environmental issues or social phenomena (Jordan and Mitchell, 2015). One of the most fundamental models within this domain is the artificial neural network (ANN), which mimics the biological neural networks of the human brain (Shanmuganathan, 2016). ANN is often considered a foundational model in the broader field of deep learning, offering a relatively simple yet effective approach to modeling.

It works by adjusting the weights between the basic recognition units, or neurons, to minimize the error between the input data and the corresponding output data. This iterative process allows the model to learn and improve its accuracy over time. When applied to spatial data, ANN proves particularly useful, as it can effectively determine the spatial relationships between data at observed locations and the various factors influencing those locations (Lee et al., 2012).

In this study, the backpropagation algorithm was employed as the key learning mechanism for the ANN with a network structure of 10 × 20 × 1. In scenarios where hidden layers exist between the input and output layers, these layers are interconnected by weight values that form a complex network of relationships. During the backpropagation process, the algorithm calculates and adjusts these weights to achieve the desired output by evaluating the slope vector of the error surface and determining the distance from the current point. This iterative process involves moving incrementally from the output layer back through the network layers, with each adjustment aimed at minimizing the overall error. By making small adjustments to the weights at each step, the model gradually reduces the error between the predicted and actual values. This process is repeated multiple times, with each iteration involving a comparison between the target and the output. The resulting error is then used to further modify the weights, ultimately improving the accuracy of the model.

ω_k0 represents the initial value of the randomly assigned weight ω_k, while ω_k refers to the updated weight for connection k. The parameter η, known as the learning rate, controls the step size during the weight update process and is typically determined through experimental tuning. E_total represents the total error value of the output, which is used to assess the discrepancy between the model’s predictions and the actual target values.

T represents the expected output value based on the input data, while j denotes the index assigned to each node within the respective layers of the neural network. n refers to the total number of output layers. The total error, E_total, is minimized through the iterative updating of the weights.

In this study, the unipolar sigmoid function was employed as the activation function, chosen for its ability to map input values to a range between 0 and 1. The equation for the unipolar sigmoid function is as follows.

In this study, for the purpose of training the ANN, forest, and non-forest locations were extracted using digital forest map data, which was constructed based on high-resolution data and field survey data, along with the NDVI images corresponding to each period. To ensure accurate classification, the top 1% of NDVI values of forest areas identified from the digital forest map were designated as forests, while the bottom 1% of NDVI values were designated as non-forest areas, with the upper and lower 0.5% of NDVI values for each period excluded. A total of 1,000 pixels were randomly selected from these areas and used as training data. The input data of the study area were normalized to values between 0.1 and 0.9 and the weights of the input layers were calculated. The initial weights were assigned randomly with the learning rate and number of epochs set to 0.01 and 1000, respectively. The root-mean-square error, which served as the stopping criterion for error reduction, was set at 0.01. Finally, based on the input data prepared in the previous steps, probability maps for forest cover were generated for both the spring and summer seasons of 2016 and 2017.

As a result, this study successfully automated the detection of forest changes by extracting self-learning data from an existing digital forest map and satellite images from each period. Using this data, a machine learning model—specifically, an ANN—was applied to generate probability maps for forest cover. In the case study, the forest probability maps were created using satellite images captured in the spring and fall, both before and after forest fires. These probability maps enabled the detection of changes in forested areas, providing a valuable tool for monitoring forest dynamics over time.

As previously mentioned, comparing the absolute values of NDVI can be challenging, even when the images are captured during similar periods (Figs. 5a, c). To address this issue, this study generated forest probability maps to quantitatively assess forest changes over time. The distribution of these probability values, as shown in Figs. 5(b, d), appears to be more normalized compared to the values derived from NDVI, allowing for a more accurate and consistent analysis of the forest.

Figure 5. Histogram.

The results of probability maps for each period are shown in Fig. 6. Notably, in the upper left region, a significant decrease in forest probability is observed following the forest fire. Throughout Fig. 6, lower probability values are evident near urban areas, particularly around roads in the central region, as well as in areas with exposed rocks, where very low values are depicted in certain sections. Additionally, factors that can affect the absolute value interpretation in a single image, such as smoke resulting from forest fires, also show reduced influence in the probability maps.

Figure 6. Result of vegetation probability maps.

The weights resulting from the training of the ANN are shown in Table 3. These weights are normalized based on the values of the Red band, which had the least influence. In the spring of 2016, the NDVI, NIR, and NDWI bands exhibited the highest influence, in that order, while in the summer, NDVI, TIR, and NIR showed the greatest impact. In contrast, for 2017, the spring image displayed a different pattern, with TIR, NDWI, NIR, and Blue bands showing the highest weights, while the summer image highlighted NDWI, TIR, and NDVI as the most influential. This analysis confirms that the NIR, TIR, NDVI, and NDWI bands consistently had the highest influence on forest probability across all periods.

Table 3 . Normalized weight of machine learning model.

Weight	20160512	20160825	20170508	20170727
R	1.0000	1.0000	1.0000	1.0000
G	1.0000	1.1886	1.4225	1.1809
B	1.2426	1.1600	1.8028	1.0851
NIR	2.1250	2.3371	1.8028	1.6383
SWIR	1.2353	1.9200	1.4085	1.4149
TIR	1.4632	2.4971	2.4648	2.1809
NDVI	2.8309	2.8514	1.5070	1.9255
NDWI	1.9265	1.7029	2.3380	2.2553
DEM	1.2868	1.7771	1.3521	1.6170
Slope	1.7647	1.8000	1.0986	1.4681

To assess the annual changes in forest cover from the forest probability maps generated in the previous step, the probability maps were paired and differentiated for each period. Figs. 7(a, b) show the differentiated images of the forest probability maps for each season, while Fig. 7c presents the cumulative annual change, derived by further differentiating the seasonal difference images. As shown in Figs. 7a and 6b, areas with no significant change exhibit values that converge towards 0 after the differentiation process such as roads and exposed rocks near the center.

Figure 7. Differential images of probability maps.

Compared to the results from the spring images, the summer images displayed a more pronounced decrease in forest probability, particularly on the eastern side. This decrease is likely attributed to the temporal difference between the two sets of summer images, which were taken approximately one month apart. When the differences from both the spring and summer images were combined, it became evident that many of the low probability values in the northwest and near the center had diminished. This indicates that by combining the changes across different periods, the effects of temporal discrepancies can be mitigated, leading to more reliable detection of forest changes over time.

Finally, to quantify the extent of forest change, the different images of forest probability generated for each period were classified using the Jenks Natural Breaks classification technique. This method allowed for the creation of forest change probability classification maps for the spring, summer, and integrated periods, respectively, providing a clearer representation of the spatial distribution and intensity of forest changes over time.

In the spring, the area with the highest grade of forest change was 851.67 ha, while in the summer it increased to 1921.95 ha, and in the integrated version, it was 723.96 ha (Fig. 8 and Table 4). Although these figures include sporadic values observed in the eastern and southern regions, the integrated version closely aligns with the 765 ha of forest fire damage reported by the Korea Forest Service for the Dogye-eup area in Samcheok-si, Gangwon-do, following the forest fire (Korea Forest Service, 2017).

Figure 8. Classification maps for forest change detection.

Table 4 . Result of forest reduction trend.

Class	Spring (ha)	Summer (ha)	Differential (ha)
High	851.67	1921.95	723.96
Moderate	9348.93	9830.16	8936.46
Low	5721.12	4169.61	6261.30

This study proposes a method for the automatic detection of forest change areas using multi-temporal, medium-resolution satellite imagery. Specifically, it presents a methodology for generating forest probability maps by period, utilizing automatically extracted forest areas based on existing datasets—such as forest maps—and NDVI values calculated from satellite images of the respective periods as training data. This approach aims to minimize the effects of atmospheric moisture, smoke, and other factors, thereby increasing the accuracy of forest change analysis. The significance of this methodology lies in its ability to mitigate the impact of atmospheric conditions and temporal differences by focusing on forest changes through the differential analysis of images captured during the same season, accounting for seasonal variations. By integrating and analyzing accumulated satellite images, this method also alleviates some of the limitations associated with optical satellite imagery, such as interference from weather conditions.

Each seasonal differential image acquired through this method, representing forest change information, can be periodically updated by overlapping the forest probability map results generated from newly acquired images for each period. This approach enables the identification of change areas over specific time frames. Additionally, by incorporating a verification system using high-resolution data such as Kompsat-3 or UAV imagery, the accuracy of forest change monitoring can be further enhanced. Through this process, continuous monitoring of forest change areas becomes feasible by overlapping and analyzing forest probability maps across multiple periods, ensuring that accurate and up-to-date forest change information is consistently provided.

In this study, as a case study applying the proposed methodology, forest changes before and after forest fires were analyzed using images from the same periods in spring and summer for the forested area of Dogyemyeon, Samcheok-si. By overlapping and differentiating images from these seasonal periods, the influence of unchanged areas was minimized, allowing for more precise detection of forest changes that reflect seasonal forest dynamics. As a result, an analysis of Landsat-8 data from 2016 and 2017 estimated the area with a high probability of reduction due to forest fires to be 723.96 ha. These findings demonstrate that forest monitoring, which accounts for seasonal variations, is an effective approach to understanding forest conditions before and after events such as forest fires.

This methodology has the potential to significantly enhance the efficiency of forest management by automating nationwide forest change monitoring through the use of CAS500-4, agricultural, and forestry satellites in the future. Given that satellite imagery of the entire country every three days, periodic data collection becomes feasible allowing for real-time monitoring of forest changes. In this process, the training set selection methodology presented in this study can be further advanced by integrating artificial intelligence alongside machine learning techniques. The automated system would not only shorten the production cycle for forest information but also facilitate the rapid identification of areas affected by forest damage, enabling prompt response measures. As a result, forest management authorities will be equipped to monitor the condition of forests across the country in real-time and implement necessary interventions more systematically and efficiently.

Finally, the automated monitoring system can contribute to the more effective implementation of targeted management and recovery measures in areas experiencing forest loss. Forest damage identified through satellite monitoring can be further analyzed using high-resolution tools such as aerial photography and UAV, enabling more precise assessments and the development of rapid recovery and restoration plans. This approach will be particularly valuable for facilitating immediate responses to forest-related disasters, such as wildfires or landslides, where real-time data provision can play a critical role in minimizing damage.

This paper was written following the research work “RE2024-04” funded by the Korea Environment Institute.

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