Burned Area Detection Using Hourly Reflectance Compositing of GOCI-II Data

Abstract

Burned area is defined as the spatial extent damaged by wildfire, where changes occur in the land surface and vegetation. This extent can be effectively observed by a satellite, which can cover large areas in a short time. In general, satellite-based burned area can be derived from temporal changes in vegetation indices such as the Normalized Difference Vegetation Index (NDVI), obtained from polar-orbiting optical satellites. However, data gaps from clouds and aerosols in optical satellite observations can make it difficult to obtain sufficient time series for burned area detection. In this study, we proposed a framework for evaluating the applicability of Geostationary Ocean Color Imager-II (GOCI-II)time series of observations, which provide higher temporal resolution than polar-orbiting satellites, for burned area detection to overcome this limitation. The framework used time series from daily compositing as the main input and time series from noon-single for comparison. Burned area was detected from both time series using a procedure adapted from the burned area mapping algorithm used to generate the MCD64A1 product and the results were compared to evaluate the effect of daily compositing when using geostationary satellite data. In commonly detected burned area, daily compositing showed a lower mean interquartile temporal range (IQR^*)than noon-single, 5.56 vs. 6.85, and higher temporal separability (S^*), 6.89 vs. 6.16. The proportion of pixels with burn-date differences within ±1 day relative to active fire detection date was also higher, 0.458 vs. 0.340.In comparison with the standard MCD64A1 product, the spatial agreement of the Dice coefficient (DC) reached 0.8534 at a burned fraction (τ) of 0.50, where the GOCI-II burned area was 107,286 ha, about 6.2% larger than the MCD64A1 product. These results suggest that high temporal resolution of GOCI-II can contribute to temporal stability in burned area detection and high spatial agreement with existing satellite-based burned area products, demonstrating the applicability of geostationary satellite data for burned area detection.

1. Introduction

Recently, wildfire trends have been increasingly influenced by human activities and climate change, resulting in substantial environmental, social, and economic impacts (Bowman et al., 2020). Wildfire causes the loss of organic matter and structural modification of vegetation, leading to substantial changes in surface properties and ecosystem structure (Keeley, 2009). Burned area can be defined as the spatial extent affected by these surface changes, and quantifying this extent is essential for monitoring the dynamics of wildfires (Chuvieco et al., 2019).

Traditionally, burned area detection has relied on field surveys and airborne surveys. Field surveys are crucial for characterizing wildfire properties and identifying ignition points, but they are difficult to apply to large wildfires, which require considerable time and cost (Chuvieco et al., 2007). Airborne surveys can provide high-resolution observations over inaccessible areas, but their application is limited by weather conditions, flight time, and the time required to collect and process data over large areas (Barmpoutis et al., 2020; Fernández-Guisuraga et al., 2018). From this perspective, satellite remote sensing has become a more practical alternative for rapidly detecting large burned areas and has been widely used to identify burned areas (Cho et al., 2023; dos Santos et al., 2023).

Satellite-based burned area detection mainly uses change detection methods based on observation pairs with pre- and post-fire spectral indices (Chuvieco et al., 2019). However, optical satellite imagery is sensitive to atmospheric conditions such as cloud cover and aerosols, which can disturb change detection based on a single observation pair. This limitation can be reduced by using high temporal resolution observations as a time series for continuous comparisons of pre- and post-fire observation pairs. In particular, this approach has the advantage of estimating the burn date at which the largest spectral change occurs within the time series in burned area detection (Giglio et al., 2018).

For this reason, global satellite-based burned area products have generally adopted polar-orbiting sensors to obtain consistent time series, while accepting coarser spatial resolution. Even with this approach, these products often rely on long temporal windows to construct time series that reduce observation gaps caused by atmospheric effects and exclude non-wildfire-related spectral signals. As these temporal windows become longer, seasonal vegetation changes can be included in the analysis, increasing the uncertainty of burned area detection (Roy et al., 2019).

In this background, geostationary satellites offer a potential way to improve the time series of observations for burned area detection (Roy et al., 2021). Geostationary satellites repeatedly observe the same region from a fixed viewing position, providing dense temporal information for compositing and statistical screening. This can help reduce data gaps and contamination caused by clouds or aerosols (Hashimoto et al., 2021). These dense time series of observations may track the timing of post-fire reflectance changes more closely and improve the stability of burned area detection.

The Geostationary Ocean Color Imager-II (GOCI-II) sensor onboard GEO-KOMPSAT-2B (GK-2B) provides Visible and Near-Infrared (VNIR) observations with improved spatial and spectral resolution and increased observation frequency compared with GOCI (Choi et al., 2021). The GOCI-II sensor has potential for burned area detection because the VNIR bands can capture post-fire changes through spectral indices such as Normalized Difference Vegetation Index (NDVI). The high observation frequency of GOCI-II also motivates an examination of whether geostationary VNIR data can improve the temporal stability of burned area detection.

In this study, we evaluate how the temporal structure of high-frequency GOCI-II time series affects burned area detection. Hourly GOCI-II reflectance data were composited into daily reflectance, and burned area was detected using a burned area detection procedure adapted from the MCD64A1 burned area detection algorithm defined by Giglio et al. (2009). Burned area results derived from time series from daily compositing were then compared with those derived from time series from noon-single to assess whether daily compositing improves the temporal stability and burned area detection.

2. Materials

2.1. Study Area

The Gyeongbuk wildfire occurred in South Korea in March 2025 under dry weather conditions and strong winds. This wildfire produced a long and narrow burned area and was reported as the largest wildfire recorded in South Korea (Ma & Kang, 2025). In this study, we focused on this single wildfire event to evaluate the applicability of geostationary satellite data for burned area detection. Therefore, the study area was defined as a smaller region within South Korea, as shown in Fig. 1(a). Land cover composition within the study area was assessed using the 10 m Global Land Cover Map 2020, as shown in Fig. 1(b). Tree cover was dominant within the study area (69.7%), indicating that the region was mostly tree-covered and suitable for assessing burned area detection based on spectral changes in vegetation.

Fig. 1. Location of the study area and land cover distribution. (a) Location of the study area (green box) within South Korea. (b) The study area for the GOCI-II burned area analysis and land cover classification based on the 10 m Global Land Cover Map 2020.

2.2. Data

2.2.1. GOCI-II

GOCI-II observes the region around the Korean Peninsula through 12 slots within a 2,500 km × 2,500 km observation domain, and each slot consists of 2,780 × 2,780 grid cells at 250 m spatial resolution. The sensor provides 12 spectral bands from 380 to 865 nm at hourly intervals for the local area (Choi et al., 2021). Among the various GOCI-II products, the Land Surface Reflectance (LSR) product provides atmospherically corrected reflectance over land surfaces using a 6SV-LUT-based atmospheric correction algorithm proposed by Lee et al. (2022). This characteristic makes it suitable for burned area detection based on pre- and post-fire spectral land changes. In this study, three months of GOCI-II LSR data from February to April 2025 were collected from Slot 7, which includes the Gyeongbuk wildfire region, to capture pre- and post-fire reflectance changes.

2.2.2. GK2A/AMI

Active fire information was obtained from the Advanced Meteorological Imager (AMI) onboard the geostationary satellite, GEO-KOMPSAT-2A (GK-2A), provided by the National Meteorological Satellite Center (NMSC). AMI is a meteorological sensor with 16 channels and provides observations every 10 min. In this study, the AMI Forest Fire (FF) product with a spatial resolution of 2 km was used. The FF product detects high-temperature anomaly pixels using thermal infrared channels and provides the location and timing of active fires (National Meteorological Satellite Center, 2024). This information is useful because active fire detections can indicate where and when burning was occurring, which supports the interpretation of reflectance-based burned area detection. FF data were collected at 10-min intervals from February to April 2025 and used as auxiliary thermal information for active fire locations and timing for comparison with the estimated burn date.

2.2.3. Data Collocation

GOCI-II LSR is observed at hourly intervals, while AMI is acquired every 10 minutes. To collocate the two datasets, each GOCI-II LSR observation was paired with the temporally nearest FF detection from AMI, with all matched pairs within a 10-minute time window. This approach was motivated by the MCD64A1 framework, in which ancillary information is linked to reflectance data. In addition, GOCI-II LSR has a spatial resolution of 250 m, while FF is provided at 2 km, so a spatial matching procedure was required to ensure consistent comparison. For this purpose, the FF data were resampled to 250 m GOCI-II pixels in the coordinate system and then matched using the nearest-neighbor method. As a result, hourly time series of active fire information were constructed. Table 1 shows the spatial and temporal resolutions of the reflectance and ancillary inputs used in the MCD64A1 product, together with their corresponding datasets in the proposed approach.

Table 1. Comparison of the satellite-based burned area algorithm components between the MCD64A1 product and the proposed approach in this study

3. Methods

3.1. Framework

The goal of this study is to evaluate how temporal input structure affects burned area detection within the MCD64A1 framework. MCD64A1 was selected because it has been widely used for global burned area mapping over long periods, allowing this study to reduce the influence of algorithm-related differences and focus on the effect of input-data structure. Fig. 2 shows the overall framework of the study. First, the GOCI-II LSR data were organized into an hourly time series and then used to generate two daily time series: one based on daily compositing and the other based on a single noon observation used as the representative daily reflectance. Next, the burned area was detected separately from these two daily time series using the adapted MCD64A1-based framework. The results from daily compositing and noon-single were compared with each other, and the result from daily compositing was further compared with the MCD64A1 product to assess spatial agreement. In the methodology, the burned area detection approach based on daily compositing is described as the main framework of this study, while the generation of the noon-single time series dataset and the reason for selecting the noon observation are described in Section 3.4.1.

Fig. 2. Flowchart of the proposed GOCI-II burned area detection and analysis.

3.2. Preprocessing

3.2.1. Valid Masking

To ensure the quality of the extracted reflectance data, three quality-masking procedures were applied. First, pixels with reflectance values outside the physical range of 0–1 were removed. These values were regarded as abnormal reflectance signals caused by atmospheric effects or observation errors. Second, quality-flag information included in the GOCI-II LSR data was used to exclude non-burnable areas such as water, so that only land pixels were retained for analysis. Third, observations acquired at times matched with active fire detection at the target pixel were excluded. This masking was intended to prevent contamination of reflectance by the radiative component of active fire (Smith & Wooster, 2005), and this mask was applied to FF matched to each pixel.

3.2.2. Cloud Masking

Optical geostationary sensors provide high temporal resolution, but surface reflectance time series can be affected by clouds, aerosols, shadows, and smoke. To reduce these contaminated observations, a masking step was applied using the UV-BLUE difference index and NDVI. While the UV band is adjacent to the VNIR spectrum, it is characterized by stronger Rayleigh scattering. This physical property effectively assists in detecting clouds and aerosols that are difficult to isolate with visible and Near-Infrared (NIR) bands alone (Jeong et al., 2016). To mitigate the risk of losing valid burned pixels due to this masking, additional criteria within a temporal window were applied. The reflectance values of UV (380 nm), BLUE (412 nm), RED (660 nm), and NIR (865 nm) bands used in this process were extracted from GOCI-II LSR. These indices were calculated as follows in Eqs. (1) and (2):

UV – BLUE difference index = UV(380 nm) – BLUE(412 nm)       (1)

\(\begin{align}N D V I = \frac {\text {NIR(865 nm)} - \text {RED(660 nm)}}{ \text {NIR(865 nm)} + \text {RED(660 nm)}}\end{align}\)       (2)

The UV-BLUE difference index was first used to identify signals unlikely to originate from the land surface. Observations with values lower than −0.05 or higher than 0.1 were masked directly. Values higher than 0.1 were also treated as unstable non-surface signals because they indicated abnormally large UV-BLUE differences. The UV-BLUE difference index thresholds were empirically determined for the dataset and the study area. These thresholds were used as conservative criteria to exclude potentially contaminated observations with extreme UV-BLUE values. For observations with UV-BLUE difference index values between −0.05 and 0.1, the original values were retained only when the NDVI conditions within a ±3-day window were satisfied. Specifically, the observation was not retained when the mean NDVI within the ±3-day window was lower than 0 or when less than 10% of observations within the window had NDVI values greater than or equal to 0.2. The NDVI threshold of 0.2 was adopted as a screening criterion based on Suárez-Fernández et al. (2023), who used NDVI > 0.2 to separate vegetated areas in wildfire dataset generation. Finally, connected land-surface candidates were labeled using 3 × 3 connectivity, and isolated components smaller than 10 pixels were removed as unstable detections. This filtering criterion was empirically applied as spatial smoothing step to reduce salt-and-pepper noise.

3.2.3. Daily Compositing

GOCI-II provides high temporal resolution but can include anomalous observations. To reduce this limitation, hourly GOCI-II LSR observations were composited to daily resolution. This approach can recover reflectance information missing by clouds and aerosols while reducing the influence of anomalous values. For each pixel, reflectance in same day were composited according to the number of valid observations. When four or more valid observations were available, the maximum and minimum values were removed and the mean of the remaining observations was calculated. When three valid observations were available, the median was used to reduce the influence of extreme values caused by residual clouds and changing atmospheric conditions. When fewer than three valid observations were available, the daily representative value was treated as invalid.

In addition, FF information was also composited to daily resolution because it was reused as an important variable afterward. To preserve fire detections within each day, a cumulative active fire mask was generated, in which a pixel was labeled as active fire if it was detected at least once on the same date. As a result, the daily time series was constructed that included both the composited daily reflectance and the cumulative daily active fire mask.

3.3. Burned Area Detection

3.3.1. Temporal Change Detection

First, RED and NIR reflectance data were extracted from daily time series, and NDVI was calculated from these bands. Then, the daily NDVI time series can be represented as consecutive adjacent window (W) pairs, where each W pair consists of a pre-fire window (W_pre) and a post-fire window (W_post). Each W is composed of valid observations and expands until the minimum number of valid observations (N_valid) is satisfied. In this study, N_valid was set to 8 following Giglio et al. (2018). A single change detection was calculated for each W pair. Then, temporal separability, S(x, y, k), was calculated as shown in Eq. (3). k denotes the observation index assigned in temporal order to daily time series and each W pair indexed by the first k included in W_pre.

\(\begin{align}S(x, y, k)=\frac{\Delta V I(x, y, k)}{\left[\sigma_{\text {pre }}(x, y, k)+\sigma_{\text {post }}(x, y, k)\right] / 2}\end{align}\)       (3)

∆VI(x, y, k) denotes the difference between the trimmed mean NDVI of W_pre and that of W_post, while σ_pre(x, y, k) and σ_post(x, y, k) denote the trimmed standard deviation of NDVI. Larger absolute values indicate more abrupt vegetation change, whereas values closer to 0 indicate gradual or no change. After S is calculated for all k, the k corresponding to the maximum S is identified as k* and S*. As a result, S* and key variables can be output as a single image representing the time series, and additional details on these variables are provided in Table 1 of Giglio et al. (2018).

3.3.2. Training Sample Selection

The pixels composing the single image representing the time series were classified into three sets based on the variables corresponding to S*. First, a priori unburned pixels were defined as those with a low likelihood of being burned and were classified when S* < 2 or σ_t* > 8. Here, σ_t* denotes a temporal texture metric representing the spatial variability of the detected change timing around each pixel (Giglio et al., 2009). Second, burned training pixels are defined as pixels likely to have burned and are identified when S* ≥ 2, σ_t* ≤ 8, and the detected change timing differed from t_f, by no more than 5 days. Here, t_f denotes the timing information derived from the cumulative active fire mask generated during the daily compositing process. Finally, unburned training samples were defined as those that were neither a priori unburned pixels nor burned training pixels and were located more than 5 km from the burned training pixels. The training sample mask is shown in Fig. 3.

Fig. 3. Training mask for the Gyeongbuk wildfire. Water and missing pixels are shown in black, unburned training pixels are shown in purple, burned training pixels are shown in red, and the remaining pixels not used for training are shown in white.

3.3.3. Probability Density Function

In this section, the Probability Density Function (PDF) was used to evaluate whether each pixel was more likely to be burned or unburned. PDF was used because temporal change detection can include false signals caused by seasonality and observation conditions; therefore, the burned and unburned likelihoods were evaluated separately instead of applying a simple threshold. ΔVI* was selected as the PDF variable because it directly represents vegetation change between the pre- and post-fire windows at the timing of maximum temporal separability. The a priori unburned and unburned training pixels were used to estimate P(∆VI* | U), whereas burned training pixels were used to estimate P(∆VI* | B).

3.3.4. Bayesian Estimation

Bayesian estimation is a statistical approach that calculates posterior probability from prior probability and likelihood (Hoff, 2009). Prior probability represents information available before considering the observation, while likelihood represents howwell the observation matches a given condition. In burned area detection, the posterior probability is defined as Eq. (4):

\(\begin{align}P\left(B \mid \Delta V I^{\star}(x, y)\right)=\frac{P\left(\Delta V I^{\star} \mid B\right) P_{B}(x, y)}{P\left(\Delta V I^{\star} \mid B\right) P_{B}(x, y)+P\left(\Delta V I^{\star} \mid U\right) P_{U}(x, y)}\end{align}\)       (4)

The prior burned probability, P_B(x, y), was defined to decrease as the distance from burned training pixels increased. It was constrained to range from 0.01 to 0.5 to prevent the prior probability from determining the burned status alone. The prior unburned probability, P_U(x, y), was calculated as 1 – P_B(x, y), ranging from 0.5 to 0.99. The likelihoods were obtained from P(∆VI* | B) and P(∆VI* | U). Through this process, each pixel was assigned a posterior burned probability. The spatial distribution of the estimated posterior burned probability is shown in Fig. 4.

Fig. 4. Posterior burned probability from Bayesian estimation for the Gyeongbuk wildfire.

3.3.5. Classification

Finally, each pixel was assigned an initial burned or unburned label based on the posterior burned probability. The initial classification was refined using nB, defined as the number of temporally consistent burned pixels among the eight adjacent neighbors. When nB < 3, an initially burned pixel was changed to unburned; when nB ≥ 6, an initially unburned pixel was changed to burned. After this refinement, the final burned area was detected.

3.4. Evaluation

3.4.1. Impact of Daily Compositing

Daily compositing is a process that combines high-temporal-resolution reflectance observations to reduce missing values in the daily time series and to obtain stable separability. To evaluate this effect, burned pixels were additionally detected from a daily time series constructed using only GOCI-II noon-single observations. The comparison between the two daily time series was designed to isolate the effect of daily compositing by analyzing burned-pixel characteristics under the same source data and algorithm structure. This evaluation focused on the effect of daily compositing on temporal stability, separability, and time difference from the AMI FF detection dates.

Noon-single observations were selected as single daily representative observations based on previous research showing that BRDF effects are reduced under lower solar zenith angles (Roy et al., 2016). The noon-single time series was constructed using the GOCI-II LSR observation acquired at 12:15 KST, which is closest to local noon among the hourly observations. When the 12:15 observation was unavailable because of quality masking or geometric observation errors, the 11:15 observation, the next closest available observation time to noon, was used instead. If both observations were unavailable, the pixel value for the day was set to invalid. Using this time series constructed in this way, an additional burned area detection result was produced.

Burned pixels detected from the two daily time series were classified into four classes according to spatial overlap and input-data condition. The four classes were used to compare variable changes at common burned locations and examine temporal and spectral characteristics of burned pixels detected only under either input condition.

1) Both-composite: burned pixels detected in both, from daily compositing.

2) Both-noon: burned pixels detected in both, from using noon-single.

3) Only-composite: burned pixels detected only in daily compositing.

4) Only-noon: burned pixels detected only in noon-single.

For each group, the interquartile range of W*(IQR*), S*, and temporal agreement between the estimated burn date and FF detection date were compared. IQR* represents the range of dates spanned by the maximum-separability window. A wider date range indicates that pre- or post-window statistics may not represent short-term snapshots of the time series, reducing the temporal stability of ΔVI* estimation. IQR* was calculated as defined in Eq. (5):

\(\begin{align}I Q R^{*}=\frac{\left(k_{(6)}+k_{(7)}\right)-\left(k_{(2)}+k_{(3)}\right)}{2}\end{align}\)       (5)

k_(n) represents the ordered observation index within W*, after the N_valid observations are arranged by time. With N_valid = 8, IQR* can be expressed as the difference between the mean of the 6th and 7th valid observation indices and the mean of the 2nd and 3rd. S* was compared with IQR* to assess whether temporal concentration of W* was associated with spectral separability, and temporal agreement with FF was used to assess whether the estimated burn date became closer to active-fire detection timing. For each group, IQR*, S*, and burn-date difference relative to FF were compared using Cumulative Distribution Function (CDF), density, and proportional bar distributions, respectively.

3.4.2. Spatial Comparison with the MCD64A1 Product

The GOCI-II burned area product derived from the time series from daily compositing was compared with the MCD64A1 product to assess spatial agreement. The MCD64A1 product is a burned area product generated from the Moderate Resolution Imaging Spectroradiometer (MODIS) daily time series of observations. The MCD64A1 product differs from the proposed method in spatial resolution and grid system, so direct pixel-level comparison may lead to scale-related bias. Therefore, 250 m GOCI-II burned/unburned binary results were reconstructed on the MODIS 500 m sinusoidal grid. When 250 m GOCI-II burned cells are aggregated to the MODIS 500 m grid, partial burns within a MODIS cell can affect spatial comparison. To examine this fraction effect, the proportion of burned GOCI-II cells within each MODIS cell was defined as the burned fraction (τ). The spatial agreement was evaluated for each τ using omission error (OE), commission error (CE), and Dice coefficient (DC) shown in Eqs. (6–8).

\(\begin{align}OE = \frac {FN}{TP+FN}\end{align}\)       (6)

\(\begin{align}CE = \frac {FP}{TP+FP}\end{align}\)       (7)

\(\begin{align}DC = \frac {2TP}{2TP + FP + FN}\end{align}\)       (8)

For each τ, reconstructed MODIS cells with a GOCI-II burned fraction greater than or equal to τ were classified as burned in the GOCI-II result. Based on this classification, true positive (TP) was defined as MODIS cells classified as burned in both GOCI-II and MCD64A1, false positive (FP) as cells classified as burned only in GOCI-II, and false negative (FN) as cells classified as burned only in the MCD64A1 product. This approach is similar to previous comparisons of medium-resolution burned area products, where partial information was considered to interpret error caused by spatial resolution differences (Tsela et al., 2014).

4. Results

4.1. Impact of Daily Compositing

In this section, burned pixels detected from daily compositing and noon-single inputs were divided into four classes based on spatial overlap, and IQR*, S*, and burn-date differences relative to FF were compared. Fig. 5 shows the spatial distribution of burned pixels detected from the two input conditions. Only three classes are shown because Both-composite and Both-noon share the same spatial locations; therefore, they were represented together as Both-all.

Fig. 5. Spatial comparison of burned area detection from daily compositing and noon-single GOCI-II observations. Cyan represents burned pixels detected by both approaches, blue represents pixels detected only from daily compositing, and green represents pixels detected only from noon-single observations.

Fig. 6(a) shows the CDF of IQR* derived for each class. The mean IQR* values of Both-composite and Both-noon were 5.56 and 6.85, respectively, indicating that the daily compositing input produced lower IQR* values at the same burned locations. This difference was also evident in the cumulative proportion at lower IQR* values, with IQR* ≤ 6 accounting for 74.15% in Both-composite and 39.82% in Both-noon. This shows that the IQR* distribution shifted to lower values in the daily compositing result. In the Only classes, the mean IQR* values of Only-composite and Only-noon were 6.45 and 7.48, indicating that Only-composite showed a lower IQR* distribution.

Fig. 6. Comparison of temporal uncertainty and spectral separability among common and method-specific burned pixels: (a) cumulative distribution of IQR*, (b) density distribution of S*, and (c) distribution of burn date relative to active fire detection.

Fig. 6(b) shows the density distributions of S* for each class to examine whether differences in IQR* were accompanied by differences in spectral separability. The S* distributions of Both-composite and Both-noon were generally similar, but the mean S* was higher in Both-composite than in Both-noon at the same burned locations, with values of 6.89 and 6.16, respectively. In the Only classes, the mean S* values were 4.38 for Only-composite and 3.85 for Only-noon, indicating that burned pixels detected only under daily compositing also showed higher S* than those detected only under noon-single input.

Fig. 6(c) shows the distribution of burn-date differences relative to FF for each class. Negative values indicate that the FF detection date was earlier than the estimated burn date. The proportions of burned pixels within ±1 day were 0.458, 0.340, 0.411, and 0.321 for Both-composite, Both-noon, Only-composite, and Only-noon, respectively. In both the Both and Only classes, daily compositing showed smaller burn-date differences relative to FF. The 0 day proportion was highest in Both-composite at 41%, more than 13 percentage points higher than Both-noon at the same burned locations, indicating closer temporal agreement with FF.

Overall, the daily compositing result showed lower mean IQR*, higher S* values, and smaller burn-date differences relative toFF in the Both classes. This indicates that, at the same burned locations, the daily compositing input produced a more temporally concentrated observation window and relatively higher spectral separability. In the Only classes, Only-composite also showed lower mean IQR*, higher S*, and closer burn-date differences relative to FF than Only-noon. In particular, Only-composite pixels were mainly distributed along the wildfire boundary,where burned and unburned surfaces can be mixed and the spectral change caused by burning may be weak. As noted by Shabanov et al. (2005), cloud and aerosol effects can further contaminate boundary signals; therefore, the higher S* and closer temporal agreement with FF in Only-composite suggest that daily compositing helped stabilize weak boundary signals.

4.2. Spatial Comparison with the MCD64A1 Product

GOCI-II burned area results derived from daily compositing were aggregated to the MODIS 500 m grid and spatially compared with the MCD64A1 product. Figs. 7(a, b) show the burned area detected by the MCD64A1 product and the GOCI-II-based result, respectively. The legend in Fig. 7(b) represents the GOCI-II burned fraction within each MODIS 500 m cell.

Fig. 7. Spatial comparison of burned area distributions on the MODIS 500 m grid. (a) Burned area detected by MCD64A1. (b) GOCI-II burned fraction aggregated to the MODIS grid. Grayscale values in (b) represent the fraction of burned GOCI-II pixels within each MODIS grid cell, ranging from 0 to 1.

As τ increased, the GOCI-II burned area decreased from 115,400 ha to 85,541 ha (Table 2). This reduction was accompanied by a decrease in CE from 0.2191 to 0.1207, while OE increased from 0.0648 to 0.2194. This opposite trend indicates that lower τ values include more partial-burn cells, whereas higher τ values retain only cells with larger burned fractions.

Table 2. Comparison of omission error (OE), commission error (CE), Dice coefficient (DC), and burned area under different GOCI-II burned fraction thresholds using MCD64A1 as the reference

DC reached its maximum value of 0.8534 at τ = 0.50, where the GOCI-II burned area was 107,286 ha, about 6.2% larger than the MCD64A1 burned area of 101,018 ha. The DC value is comparable to the 0.8620 reported by Stroppiana et al. (2025) for burned-area comparison across different spatial resolutions. At τ = 0.75, the area difference became smaller, but DC decreased to 0.8510, indicating that area similarity and spatial agreement are not necessarily maximized at the same burned fraction.

The trade-off between CE and OE is consistent with previous studies showing that mixed pixels and spatial-resolution differences can be major sources of spatial disagreement in burned area comparisons (Franquesa et al., 2022). Therefore, the highest spatial agreement at τ = 0.50 suggests that MCD64A1 correspondence was better explained when partial-burn cells were included to some extent, rather than when only cells with high burned fractions were selected. This also suggests that the GOCI-II result captured the main burned pattern of MCD64A1 while providing a more detailed representation of boundary areas and partial-burn cells at 250 m spatial resolution.

5. Discussion

The results indicate two main findings. Daily compositing improved the temporal stability of the GOCI-II time series compared with the noon-single time series, producing more stable windows and closer agreement with FF detection timing. The GOCI-II burned area result was generally consistent with MCD64A1 in spatial pattern and overall burned area scale, while the burned fraction provided additional detail for the boundary. However, these results should be interpreted with several considerations.

First, the difference in spatial resolution between the GOCI-II LSR product (250 m) and the AMI-FF product (2 km) should be considered when interpreting the results. In this study, the AMI-FF product was used to mask reflectance observations, while the final burned area detection was based on temporal changes in the GOCI-II LSR time series. Therefore, the spatial resolution of the AMI-FF product did not directly determine burned and unburned pixels, but it could affect the availability of valid observations before daily compositing. This effect is expected to be limited in the central burned area, where the temporal change signal is relatively strong. However, this uncertainty may be larger near burn boundaries, where surface conditions are spatially heterogeneous and burned signals are weaker. This uncertainty may affect the detection of boundary burned pixels, and future work could further assess this effect.

Second, the temporal accuracy of the detected burn date should be interpreted with caution. In this study, comparison of burn date with the AMI-FF product was used to examine the temporal consistency with active fire timing and evaluate the relative effects of daily compositing within the same GOCI-II-based framework. The AMI-FF product provided valuable hourly baseline information for this purpose. As AMI-FF was also used in the processing, this comparison should be interpreted as an internal comparison rather than as validation against a completely independent reference dataset for the absolute temporal accuracy of the detected burn date. Future work should include comparison with independent reference data to evaluate the temporal accuracy.

6. Conclusions

This study evaluated burned area detection using high-temporal-resolution GOCI-II reflectance data for the 2025 Gyeongbuk wildfire. In the same algorithm structure, daily compositing showed lower temporal variability, higher separability values, and smaller burn-date differences relative to FF than noon-single input. These results indicate that daily compositing produced more stable temporal windows and closer temporal agreement with active fire detections. These results suggest that high-temporal-resolution VNIR observations from GOCI-II can provide complementary information for burned area detection and temporal uncertainty reduction. Since this study was conducted as a case study using an MCD64A1-based framework, further comparisons with other burned area detection algorithms are needed to identify structures more suitable for high-temporal-resolution data. Future work should also extend toward new burned area detection algorithms that better use geostationary temporal observations while considering spatial agreement, temporal stability, and shorter production intervals.

Author Contributions

Conceptualization: Jung HC, Hwang D-H, Kim SJ, Shin KH; Data curation: Choi J-K, Lee K-S; Formal analysis: Jung HC, Shin KH; Funding acquisition: Jung HC, Hwang D-H; Investigation, Methodology: Jung HC, Hwang D-H, Kim SJ, Shin KH; Project administration, Supervision: Jung HC; Validation: All authors; Writing–original draft: Shin KH; Writing–review & editing: All authors.

Conflicts of Interest

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

Funding

This research was supported by the National Research Foundation of Korea (NRF) grant funded by the South Korea Government (RS-2026-25483027), and Global-Learning & Academic Research Institution for Master’s, PhD Students, and Postdocs (LAMP) program of the National Research Foundation of Korea (NRF) grant funded by the Ministry of Education (RS-2024-00442483), and the Rural Development Administration (RDA) grant funded by the Ministry of Agriculture, Food and Rural Affairs(MAFRA)(RS-2026-25515531).

Data Availability Statement

Data sharing is not applicable to this article as no new data were created or analyzed in this study.