A Comparative Study Between Gaofen-1 WFV and Sentinel MSI Imagery for Fire Severity Assessment in a Karst Region, China

Abstract

Wild fires frequently influence fragile karst forest ecosystems in southwestern China. We evaluated the potential of Gaofen Wide Field of View (WFV) imagery for assessing the fire severity of karst forest fires. Comparison with Sentinel Multispectral Imager (MSI) imagery was conducted using 19 spectral indices. The highest correlation for Sentinel-2 MSI is 0.634, while for Gaofen-1 WFV it is 0.583. This is not a significant difference. The burned area index, differenced burned area index, and relative differenced modified soil adjusted vegetation index were the highest performing indices for the Gaofen-1 WFV, while the normalized burn ratio plus, differenced normalized differential vegetation index, and relative differenced normalized differential vegetation index were the best for the Sentinel MSI. The total accuracy evaluation of the fire severity assessment for Gaofen-1 WFV ranged from 40 to 44% and that for Sentinel MSI ranged from 40 to 48%. The difference in accuracy between the two satellites was less than 10%. The RMSE values for all six models were close to 0.6, ranging from 0.58 to 0.67. The fire severity maps derived from both imagery sources exhibited overall similar spatial patterns, but the Sentinel-2 MSI maps are obviously finer. These maps matched well with the unmanned aerial vehicle (UAV) images, particularly at high and unburned severity levels. The results of this study revealed that the performance of the Gaofen WFV imagery was close to that of Sentinel MSI imagery which makes it an effective data source for fire severity assessment in this region.

1. Introduction

Every year, vast expanses of forestry covering millions of hectares worldwide experience severe destruction due to forest fires, and large-scale fires frequently occurring in recent years have emitted significant amounts of carbon dioxide (CO₂) [1]. In addition to the emission of CO₂, satellite observations have detected a significant increase in another toxic gas, CO, during the wildfire events [2]. Forest fires serve significant functions within the environment, ecosystem [3,4], and climate [5], as well as affecting human lives and property [6]. Between 1987 and 2007, China witnessed roughly 8200 forest fires yearly, with an average annual burned area of approximately 400,000 hectares, causing severe damage [7]. More recently, between 2005 and 2018, a total of 122,952 fires occurred in China [8], averaging 8782 fires each year, which slightly increased compared to the period from 1987 to 2007. Among the various factors causing forest fires, temperature and humidity play crucial roles [9], and as global warming continues, more forest fires are expected to occur. Forest fires display clear spatial characteristics over China, with most occurring in southern and northeastern regions [8]. The fire season in the subtropical evergreen region of China is long, spanning from November to May [10,11]. Climate statistics indicate that wildfires in this region are notably influenced by a periodic climate fluctuation known as El Niño—a condition where unusually warm ocean temperatures in the equatorial Pacific Ocean disrupt normal weather patterns, particularly during notable years such as 2010 and 2014 [8,12]. The frequency of El Niño may increase due to climate change [13], which may also predict an increase in forest fires in the future.

Many factors can influence the reforestation process, including the length of time after fire and fire severity [14]. Since the end of the last century, fire severity—defined as the intensity and heat release of fires—has been assessed using various satellite-based spectral indices [6,15,16,17]. These fire severity indices serve as an important means of providing post-fire restoration information [18,19,20,21]. The difference in the normalized differential vegetation index (NDVI) between pre- and post-fire images was used to assess fire severity in Australia in 2004, employing a high-resolution SPOT2 satellite [22]. The NDVI demonstrated strong performance, achieving a classification accuracy of 88%; the main biases were observed in the middle and low severity classes. In 2006, a new normalized burn index, the normalized burn ratio (NBR), was systematically applied to assess fire severity [23]. Using index theory analysis and land investigations conducted at various locations, the study meticulously evaluated the ability of the NBR to depict fire severity. Analysis using index theory suggested that the responsiveness of the NBR to fire intensity could vary, indicating a lack of consistent sensitivity. An analysis of MODIS and Landsat ETM+ data for Australian savannas, boreal forests in the Russian Federation, and South American tropical forests revealed that shortly after fire events, the NBR (normalized burn ratio) is not likely the most suitable spectral index for describing fire severity [24]. Although NBR is theoretically not an ideal spectral index to depict fire severity, the correlation can range from 0.59 to 0.83 between pre- minus post-NBR (dNBR) and the composite burn index (CBI) acquired through land surveys [25,26], which indicates that the NBR can effectively evaluate fire severity. The NBR was also recognized as a stable fire severity index after it was calculated using 13 spectral indices and correlated with fire severity estimates derived from field surveys of four wildfire burn sites in Alaska. Notably, in this study, regardless of single-date post-fire or bi-temporal analyses, the NBR could accurately estimate fire severity within forested areas [25]. Furthermore, a study of three fires in southern Spain indicated that dNBR is more effective at differentiating between unburned and burned pixels, while NBR is appropriate for distinguishing between extreme and moderate severity levels [27]. A relatively strong correlation of 0.84 between the dNBR and CBI was also observed for a fire in San Francisco, Arizona, USA [28]. Many studies have indicated that changes in the NBR are consistent with land surface surveys of different forest types in the southwestern USA [28,29]. A new version of the NBR, the relative version of the dNBR (RdNBR), was developed based on examining observational data gathered from 14 wildfires in California, USA. The RdNBR study aimed to reduce misclassification errors in low-vegetation areas or when field data were unavailable [30]. Although RdNBR can enhance the classification precision of high severity locations, there was a decrease in the precision of unchanged and low severity areas. Therefore, the overall classification precision was equivalent to the dNBR [30]. In 2014, an index combined the dNBR and NBR, known as the relativized burn ratio (RBR), as defined by Parks et al. [20]. This ratio was proposed to address some denominator difficulties. The results of this investigation indicated a slightly better result over the western USA than both the dNBR and RdNBR [20]. A study conducted in central Chile compared the performance of dNBR, RdNBR, and RBR in evaluating fire severity, and the results indicated that RdNBR outperformed the other two indices [31]. Consequently, considering the location and physical situation of the burned area is vital when deciding whether to use the dNBR or RdNBR in mapping fire severity, or the dNDVI and other transformation indices.

Apart from medium-spatial-resolution satellite imagery, various high-spatial-resolution images are available for fire severity assessment, such as IKONOS, Sentinel-2, and air platform remote sensing. Pixel- and object-based classification methods are available for high-resolution imagery. Pixel-based classification methods only use the spectral information of the images, similar to that used for the NBR, dNBR, NDVI, and dNDVI. The object-based classification method has been used in fire severity mapping of high-resolution spatial imagery, using both spectral and contextual information. This method proved an effective methodology for fire severity classification on Thasos, a northern Greek island, achieving an overall accuracy of 83% and a Kappa coefficient of 0.74 [32]. Unlike pixel-based methods, object-based methods use object information consisting of a group of pixels that form meaningful objects or features. Notably, using a combination of pixel- and object-level approaches for evaluating fire severity achieved better results than any other approach. The combination method acquired an overall accuracy of 97.32% for a wildfire in Spain, demonstrating the benefits of using both spectral and contextual information [33]. For a Mediterranean forest ecosystem, post-fire land surface temperature (LST) was a potentially valuable variable for fire severity assessment [34]; however, LST or its delta pattern is not effective in every scenario [21]. Subsequently, an index combining LST and the enhanced vegetation index (EVI) was proposed to solve the insufficient representation of fire severity in areas with unburned trees and vegetation [35]. Previous studies proposed a support vector regression (SVR)-based method and random forest model can achieve more accurate results than the regression method based on the dNBR [15]. The semi-supervised transfer component analysis-based support vector (SSTCA-SVR) model was proposed after the development of the SVR model to address the challenge of assessing burned areas where field data are unavailable. This model incorporates semi-supervised transfer component analysis into SVR, demonstrating superior performance while achieving higher correlations and lower errors compared to dNDVI, dLST, dNBR, and the traditional SVR model [21]. The machine learning model, XGBoost, which is based on the NDVI, NBR, burn index, and other auxiliary factors in Australia, achieved an accuracy of over 86% when using the dNBR as the validation dataset [16]. As a complication of forest fires, each method was tested in a finite number of scenarios, and the choice of which model to use should be based on the availability of field data and the type of forest.

Karst areas are uniquely distributed globally and occupy 15% of the earth’s surface [36], playing a vital role on our planet. With the fragile ecosystem, protecting the environment in this region faces many challenges, especially in the largest Southeast Asia karst region [37]. The karst area in southwestern China offers a range of ecological benefits, particularly in terms of carbon storage; however, forest fires constantly pose a threat, as this region is one of the major forest fire regions in China [38]. Therefore, conducting research on forest fire severity plays a crucial role in assessing the regional impact of forest fires and evaluating forest recovery. Guizhou province, which is a key part of the karst area in southwestern China, is an ideal place to carry out relevant research. From 2023 to mid-February 2024, Guizhou Province experienced a burst of fires, which presents a good opportunity to analyze fire severity in this region. Unfortunately, the period from 2023 to 2024 was also marked by a strong El Niño event, which may once again prove the correlation between El Niño and forest fires in this area.

Guizhou is located in the region with the lowest sunshine duration region in China, where the annual sunshine hours are around 1500, leading to a shortage of optical satellite data resources. Gaofen is a series of Chinese satellite platforms that offer high spatial resolution, as well as capabilities in hyperspectral and SAR (synthetic aperture radar) imaging. Additionally, some Gaofen satellites, such as those equipped with WFV on Gaofen-1 and Gaofen-6, also provide high temporal resolution; many forest fires have been covered due to its short return period. In 2019, the China National Space Administration (CNSA) announced free access to 16 m resolution Gaofen WFV instrument data from the Gaofen-1 and Gaofen-6 satellites for the world. This provides an abundant option for modern high-resolution imagery to evaluate fire severity. Most of the studies on fire severity discussed above were based on the use of the Landsat, SPOT, and Sentinel satellite imagery. With the open access to Chinese Gaofen satellite imagery, obtaining data collected using the WFV camera on the Gaofen satellite is convenient. There are limited studies that have used both Gaofen-6 WFV and Gaofen-1 WFV images to extract burned areas using the dNDVI or Otsu’s method, with high extraction accuracy reported [39,40]. However, assessments of fire severity using Gaofen satellites are rare. To evaluate fire severity of forest fires in karst forests using Gaofen satellites, it is necessary to assess their applicability in this region. Sentinel-2 MSI images achieved a better overall performance in fire severity evaluation than that of Landsat 8 OLI images for spatially heterogeneous Mediterranean forest ecosystems [41,42,43]. Given the comparable spatial resolution of Gaofen WFV imagery, it is appropriate to choose Sentinel-2 MSI as a cross-comparison source. The significance of comparing Gaofen satellites with Sentinel-2 satellites lies in the need for a comprehensive evaluation of their capabilities, especially in the context of assessing forest fire severity in karst regions, as such an evaluation in this kind of forest is rare. While ground-based evaluations of sampling points offer definitive data, the unique and complex nature of karst forests necessitates a horizontal comparison to complement these findings. This is where Sentinel-2 satellite data become invaluable, providing a basis for cross-comparison that enables a more scientific and accurate assessment of Gaofen satellites’ performance in this specific application. By conducting such a comparison, we can determine whether the Gaofen satellite is capable of assessing fire severity in karst forests, ultimately enhancing our ability to monitor and evaluate forest fires in karst ecosystems. Thus, we evaluated the performance of Gaofen-1 WFV imagery for determining fire severity levels (compared with those of the Sentinel-2 spectral indices), as well as its differential forms and relative versions, to determine the adaptation of Gaofen-1 data in fire severity evaluation. Therefore, we first compared the spectral difference between the Gaofen-1 WFV and Sentinel MSI instruments in this study, and their performance on fire severity assessment in a karst forest, based on a forest fire that occurred in 2023. Subsequently, we introduced three indices rarely used in severity assessments: the global environmental monitoring index (GEMI), BAI, and NBR+. These indices have performed well in burned area identification. Further, we utilized widely used indices, such as the NDVI, modified soil adjusted vegetation index (MSAVI), and NBR. The main objective of this study has three aspects: (1) based on field-collected CBI data, this study aims to identify which indices derived from Gaofen satellite are capable of assessing fire severity in the heterogeneous karst forests; (2) to determine the optimal model for evaluating fire severity in karst region forests utilizing Gaofen-1 WFV imagery; and (3) to compare the performance in fire severity assessment with Sentinel MSI imagery to cross-validate the performance of Gaofen-1 WFV.

2. Materials and Methods

2.1. Fire Events

In this study, we focused on a forest fire that occurred on 3 April 2023, in Wanshui Town, Guizhou Province, in the southwestern region of China (Figure 1a). This was a typical karst forest fire that burned an area of 580,186 m² (Figure 1b), contrasting with most fire severity studies that have primarily focused on large-scale burned areas. Instead, due to the patchy landform characteristic of the karst region and human disturbances, many burned areas in this region are typical of medium to small scales and exhibit a dispersed distribution, as exemplified by this particular fire. Fires in this area tend to affect smaller areas due to the local government’s robust fire management strategies. This area belongs to Kaili City, Guizhou Province, with an average annual precipitation and temperature of approximately 110 cm and 16.6 °C [44], respectively. The elevation varies from 529 to 1447 m, revealing dramatic relief. The eastern side of the mountain is very steep, with the steepest slope reaching 68° (Figure 1c,d). Additionally, the land cover is dominated by shrub-herb communities, shrubbery, and trees. The main vegetation is Pinus tabuliformis Carrière (Chinese pine), Castanea mollissima Blume (Chinese chestnut), and Quercus alba L. (white oak), followed by Cunninghamia lanceolata (Lamb.) Hook. (China fir), ferns, Dicranopteris dichotoma (Thunb.) Bernh., Imperata cylindrica (L.) Raeusch., and others.

2.2. Imagery and Preprocessing

Landsat images have been used in fire severity assessments for decades because Landsat was the only free medium-high spatial resolution remote sensing dataset before 2018. Subsequently, the European Space Agency (ESA) released all Sentinel images. The Sentinel-2 MSI sensor on board Sentinel-2A and Sentinel-2B satellites possesses 13 reflective bands, from the VIS to the SWIR band, with a spatial resolution range of 10–60 m, which can be acquired through the Copernicus website [45,46]. The Gaofen-1 WFV sensor has four Vis-NIR reflective bands with 16 m spatial resolution [47]. Figure 2 shows a comparison of the spectral bands from the Gaofen-1 and Sentinel-2B satellites. We collected pre-fire imagery from Gaofen-1 and Sentinel-2B in November 2022; post-fire imagery was also obtained close to the timing of the field survey campaign to minimize the influence of plant phenology, solar angle, and atmospheric situation (Table 1; [48,49]). Although Sentinel-2 has L2A data available, which has been subjected to atmospheric correction, these data appeared strange in some regions, especially on the sunless side of the mountain; therefore, we downloaded the L1C data and processed it to L2A data using the SEN2COR toolbox from ESA. After correcting for atmospheric effects, we used the SEN2RES tool to super-resolve the image to a 10 m resolution for all of the lower-resolution bands.

For Gaofen-1 WFV, data close to the field survey date were collected from the CNSA data platform (Table 1). The DN values of the Gaofen-1 WFV were transformed into the top of atmospheric radiance based on gains and bias coefficients from the CNSA website [47]. We employed the MODTRAN model from ENVI 5.3 within an atmospheric correction tool to account for atmospheric effects and obtained the ground reflectance. Furthermore, an orthorectification correction procedure from ENVI 5.3 (Harris Geospatial Solutions, Melbourne, FL, USA) was employed to correct the geometric influences to produce an orthorectified image. There was some geolocation bias in the image, so a manual image registration procedure was performed, where the Sentinel MSI image was used as the base image in the ArcGIS 10.1 software (EsriChina, Beijing, China).

2.3. Fire Severity Spectral Indices

Considering the spectral band distribution of Gaofen-1 WFV and Sentinel MSI, among the numerous spectral parameters reported by various studies, nineteen reflective indices were selected (Table 2). These reflective indices were classified into three categories: spectral indices, differenced spectral indices, and relatively differenced spectral indices. NBR, NBR+, their differenced and relatively differenced forms, and RBR, all of which require the SWIR band, were calculated for Sentinel MSI images and not for Gaofen-1 WFV images, whereas the remaining indices were computed for both of them.

In calculating the NDVI, BAI, and MSAVI for Sentinel MSI images, Band8 was selected for the NIR band over Band8a due to its spectral proximity to Gaofen-1 WFV bands (Figure 2).

2.4. Field Survey Data

The CBI is a widely used and effective in situ metric of fire severity; within the initial vegetation growing season following the occurrence of a fire, a total of 75 plots (16 × 16 m) of field data were acquired (Figure 3; detailed field samples in Appendix A). For each plot, using the method of Key and Benson [23], up to five strata of fire severity were visually investigated and scored from zero to three in each stratum. The five strata were divided into two categories: the understory and the overstory. The understory consisted of three strata: the substrates (herbs, low shrubs, and trees less than 1 m tall), tall shrubs, and trees 3–16 ft tall. Considering that the duff layer above the soil is relatively thin in the karst region, the duff substrata were not included in the substrate strata. In the overstory, which consisted of the substrata of intermediate and large trees, we increased the tree mortality rate, which can be visually observed more easily. An ordinary averaging process was applied to the first three levels, the last two levels, and all five levels, in separate analyses. Sample plots were selected based on accessibility and preliminary severity assessments derived from post-fire imagery. Figure 4 shows the photographs of the field sampling sites.

To establish the correlation between remote sensing spectral indices and field-collected CBI data, the GPS coordinates recorded on-site during data collection were matched with their corresponding pixels in the satellite images. To mitigate the impact of GPS positioning errors and geolocation bias in the satellite images, the spectral indices from the satellite images in the sample plot were subsequently computed by applying a 3 × 3 pixel mean filter algorithm. The matching procedures were implemented using the ArcGIS 10.1 software on a Windows 10 workstation (ESRI China, Beijing, China). After the spectral indices were extracted and averaged within the 3 × 3-pixel matrix, rigorous statistical analyses were carried out to evaluate the correlation between remote sensing spectral indices and field-collected CBI data.

2.5. Statistical Analyses

The least squares method was used in the data analysis, as it is commonly employed in similar studies [30,42]. The Pearson correlation coefficients between spectral indices and field-collected CBI were computed for the entire dataset to assess the statistical significance of their correlations for each index. The performance of the indices was compared based on the correlation coefficients. The determination coefficients (R²) for the three best models from each satellite were calculated and presented using scatter plots for both the entire dataset and the validation dataset. After the linear models for fire severity assessment were established on the training dataset using the least squares method, the root mean square error (RMSE) was calculated on the validation dataset as follows:

(1)$RMSE=\sqrt{\frac{1}{\mathrm{n}}{\in }_{\mathrm{i}=1}^{\mathrm{n}}{({\mathrm{y}}_{\mathrm{i}}-{\mathrm{y}}_{\mathrm{i}\mathrm{p}})}^2}$

where n represents the number of the validation dataset, ${\mathrm{y}}_{\mathrm{i}}$ denotes the field-collected CBI, and ${\mathrm{y}}_{\mathrm{i}\mathrm{p}}$ indicates the model-predicted CBI.

Thresholds of three outperformed spectral indices for different fire severity levels were established using the linear models developed; the classification accuracy was calculated on the validation dataset as follows:

(2)$Accuracy={\displaystyle \frac{{\mathrm{N}}_{\mathrm{j}\mathrm{p}}}{{\mathrm{N}}_{\mathrm{j}}}}\times 100\mathrm{\%}$

where ${\mathrm{N}}_{\mathrm{j}\mathrm{p}}$ is the number correctly predicted by the model and ${\mathrm{N}}_{\mathrm{j}}$ is the total in that class. Figure 5 presents a comprehensive flowchart outlining the methodological steps employed in this study.

3. Results and Discussion

3.1. The Comparative Analysis of Different Indices in Fire Severity Assessment

Table 3 presents the Pearson correlation coefficients for Gaofen-1 WFV, Sentinel MSI, and CBI obtained from the correlation analysis. All spectral indices, except for NDVI, exhibited statistically significant correlations (p < 0.001) with the field-collected CBI. For Sentinel MSI imagery, NBR+, dNDVI, and RdNDVI were the most strongly correlated spectral indices within their respective categories, whereas in Gaofen-1 WFV data, BAI, dBAI, and RdMSAVI demonstrated better performance.

In Sentinel-MSI, NBR+ outperformed NBR, likely due to its inclusion of blue-green band information and a broader value range [55]. Although RdNDVI had the highest correlation coefficient, the difference between dNDVI and RdNDVI was small (0.625 vs. 0.635), suggesting that the differenced index also performed well in fire severity assessment. The dNDVI showed similar correlation patterns to NBR-based indices (dNBR and RdNBR) in the Mediterranean ecosystems of Spain [42], as well as within a Greek Mediterranean pine ecosystem [41]. The use of differenced and relatively differenced vegetation indices for NBR has been widely debated; no conclusive agreement has been reached on which version is better with field-based fire severity data [20,56]. While the performance of the NBR lagged slightly behind that of NBR+ in the single index category, it still proved effective in fire severity assessment, as demonstrated in previous studies [25]. The minimal correlation difference between NBR+ and NBR and the even smaller difference between dNBR+ and dNBR illustrate the effectiveness and similarity of these two indices.

For Gaofen-1 WFV, the BAI showed strong performance, probably due to its superiority over NDVI, SAVI, and GEMI in distinguishing burned areas [51]. This highlights its importance in detecting fire-related signals. BAI, dBAI, and RdBAI performed admirably across all indices, with correlation coefficients exceeding 0.5, while GEMI demonstrated comparable performance. In burned area detection using Gaofen-1, BAI and GEMI surpassed NDVI and EVI, indicating that these two indices also have the potential for assessing fire severity [57]. The relative version index-RdMSAVI exhibited the best performance among the three indices in all categories. This is likely due to its emphasis on relative changes in pre-fire vegetation cover, which results in comparatively different vegetation values despite levels of fire-induced vegetation mortality [30]. This indicates that in areas with low vegetation cover before the fire, where all vegetation is consumed in a high severity fire, the difference in vegetation cover will yield a low value. The induced denominator in the relative difference index equation results in a more sensitive response compared to the difference indices under low vegetation canopy conditions before fire [20]. The scatter plots of the best performing indices from both Gaofen-1 WFV (Figure 6a,c,e) and Sentinel MSI (Figure 6b,d,f) revealed the existence of a linear correlation between these spectral indices and CBI. The regression model exhibited a high degree of fit to the data, indicating a significant trend between the CBI and remote sensing indices, which was observed when using Gaofen-1 WFV and Sentinel MSI data.

3.2. The Fire Severity Model Construction

To map fire severity, determining the thresholds for different fire severity classes is essential. Different classification thresholds have been proposed for fire severity assessment. In Key and Benson (2006) [23], the defined ranges for severity levels were as follows: zero for unburned areas, [0.5, 1] for low severity, [1.5, 2] for moderate severity, and [2.5, 3] for high severity. Mitri and Gitas (2008) [32] developed three fire severity levels: CBI > 1.7 for heavily burned areas, 1.0 ≤ CBI ≤ 1.7 for moderately burned areas, and CBI < 1.0 for slightly burned areas. Meanwhile, Miller and Thode (2007) [30] used the halfway values between CBI classes to classify severity into four distinct levels: unchanged, low, moderate, and high. The choice of which CBI threshold to use depends on factors such as value judgment, management objectives, analysis criteria, and others [30]. In this study, we adopted the halfway method to split the CBI dataset into four severity levels, specifically assigning values greater than or equal to 2.0 (CBI ≥ 2.0) to the high severity category, as no field CBI values exceeded 2.25. Based on this CBI threshold, a sample dataset of 75 plots was randomly split, with 50 plots forming the training dataset and 25 plots comprising the validation dataset (Table 4). Using the best performing indices from each category, fire severity assessment models for Gaofen-1 WFV and Sentinel MSI were established using the training dataset (Table 5).

Although the determination coefficients of the models derived from the Sentinel-2 MSI were higher overall, the models based on Gaofen-1 WFV showed slightly better performance in predicting CBI on the validation datasets (Figure 7). The R² values for the RdMASVI and dBAI models from Gaofen-1 WFV were 0.393 (RMSE = 0.579) (Figure 7a) and 0.345 (RMSE = 0.605) (Figure 7c), respectively, whereas the R² values for the RdNDVI and dNDVI models from the Sentinel MSI were 0.377 (RMSE = 0.594) (Figure 7b) and 0.350 (RMSE = 0.613) (Figure 7d), respectively. On the validation dataset, for Gaofen-1 WFV, the RdMSAVI model performed better than the dBAI model. Similarly, the RdNDVI models from Sentinel MSI also performed better than the dNDVI model. Finally, the BAI model derived from Gaofen-1 WFV and the NBR+ model from Sentinel-2 MSI performed the poorest (Figure 7e,f).

The model thresholds were determined based on the CBI thresholds of different fire severity and the model formulas from Table 5 (Table 6). Employing the defined thresholds, we determined the precision of each model in assessing fire severity on the validation datasets (Table 7). The RdMSAVI model outperformed the BAI and dBAI models in Gaofen-1 WFV imagery, whereas RdNDVI performed better for Sentinel MSI data. Therefore, the RdMSAVI model from the Gaofen-1 WFV was selected for mapping the fire severity at this site and compared with the RdNDVI model for Sentinel MSI. The thresholds for the RdMSAVI model from Gaofen-1 WFV, calculated from the CBI thresholds, were as follows: unburned (RdMSAVI ≤ −49.75), low (−49.75 < RdMSAVI < 26.08), middle (26.08 ≤ RdMSAVI < 82.95), and high (RdMSAVI ≥ 82.95). For Sentinel MSI, the thresholds for the RdNDVI model were unburned (RdNDVI ≤ −117.64), low (−117.64 < RdNDVI < −61.64), middle (−61.64 ≤ RdNDVI < −19.63), and high (RdNDVI ≥ −19.63) (Table 6).

3.3. Fire Severity Mapping

The fire severity map derived from the Gaofen-1 RdMSAVI model exhibited a comparable spatial pattern to that derived from the Sentinel-2 MSI RdNDVI model (Figure 8). Regarding spatial distribution, the high severity class was mainly surrounded by the middle severity class, which showed a reasonable distribution pattern (Figure 8a,b). Based on land investigation and UAV imagery, its location also matched reality (Figure 8c,d), especially the high and low severity levels. The fire severity map derived from Sentinel-2 MSI imagery indicated a more mosaic pattern, potentially due to the imagery’s higher resolution or the highly complex environment. In the Gaofen-1 RdMSAVI model, the overall statistics of fire severity indicate that the unburned class (pixels) represented the smallest proportion of the burned area, while the fire severity of the high class accounted for the largest area (Figure 8e). For the Sentinel-2 MSI RdNDVI model, the low severity class accounted for the largest proportion while the unburned class remained the smallest (Figure 8f). Although there were differences in the proportions of these two models, the variation was within 5%. The results indicate that the Gaofen-1 WFV imagery provided a comparable fire severity assessment to the Sentinel-2 MSI imagery.

In the Gaofen-1 WFV-derived fire severity map, a few isolated pixels were located centrally within the high severity area. Upon inspection, we found that their values were only slightly below the threshold for the high- severity class. As a result, we slightly adjusted the threshold for high severity from 82.95 to 81.48. Although some isolated pixels still appeared in the high severity class, they now occur at the edges, which seems reasonable. After this slight adjustment, the fire severity map for the site was redrawn (Figure 9).

3.4. Discussion

This study presented the evaluation of Gaofen-1 WFV imagery in fire severity assessment in a karst forest fire. Among all of the collected CBI datasets, although the Gaofen-1 WFV images exhibit slightly poorer performance compared to Sentinel-2 MSI images, their correlation values remain close, with the highest correlation reaching 0.583 among all indices, versus 0.634 for Sentinel-2 MSI image. Although the correlation indices in this study (r = 0.634) are relatively lower than those reported in previous studies—for instance, a dNDVI-based correlation of 0.779 was achieved in a forest fire analysis in northwestern Spain using Sentinel-2 MSI data [42]—they nevertheless show statistically significant correlations with field-collected CBI (p < 0.001). In the severity map, when compared to Gaofen-1 WFV, Sentinel-2 MSI provided a finer result, although the overall distribution is similar. The superior performance of Sentinel-2 MSI imagery can be attributed to its higher spatial resolution (10–60 m) and specialized spectral bands, particularly in the red-edge region. Multiple studies have demonstrated this advantage [41,42]. However, the difference in the proportion of different fire severity levels between the two satellites is less than 5%. The overall results indicate that Gaofen-1 WFV imagery is also an effective satellite resource in assessing fire severity in this karst forest region, despite Sentinel-2 MSI imagery yielding finer results likely due to its higher resolution. The preliminary use of UAV images in this study relies solely on qualitative analysis; it uses the UAV image to select sample points by visual interpretation [40], but further research is needed to better incorporate the remaining information in the UAV data.

Among the 19 indices across the three categories, NDVI was the least correlated index for both Gaofen-1 WFV and Sentinel-2 MSI imagery, consistent with findings from previous studies [42,58]. This may be attributed to the favorable water and heat conditions in the subtropical region, which promote faster vegetation growth after fire, especially in places with higher severity. As a result, the NDVI values across various locations have become more similar. The better performance of RdNDVI and dNDVI for Sentinel-2 MSI in this region may suggest that, in karst regions like Guizhou, the lack of the SWIR band does not influence their capability in fire severity assessment. But in similar research studying the impact of a forest fire in the Mediterranean biogeographic region [42], dNBR and RdNBR were found to perform better in assessing fire severity, followed by dNDVI as the next most effective index for Sentinel-2 MSI imagery. However, comparing results from previous studies with those of this study is difficult, as the fire severity assessments can vary significantly across different ecosystems and among different investigators. Therefore, further research should be conducted to confirm this finding.

The overall accuracy is relatively lower than some similar studies. In the work on a forest fire in the Mediterranean region of Türkiye, the dNBR achieved more than 80% overall accuracy [2]. And in the work on a forest fire in a Mediterranean pine ecosystem of Greece, the overall accuracy reached 73% [41]. However, in comparison to some other studies, the result is better. Specifically, in a forest fire that happened in the Vesuvius National Park in Italy, the overall accuracy of the Sentinel-2 dNBR model was 30% [43]. Some limitations can be found for this. One limitation is the heterogeneity of the karst forest. At this forest fire site, there is grassland, shrub land, and forest land, making the field investigation difficult. A visual survey may not fully capture the characteristics of the sample site. This may also introduce some bias in representing the real fire severity level at the sample site. As there are some problems with the positioning of Gaofen-1 WFV imagery, although careful corrections have been made, some degree of bias still exists. This results in a less than ideal match between spectral indices and field CBI in some sample sites. Given the highly heterogeneous surface conditions, although a 9 × 9 average was calculated, this may lead to a decrease in the representativeness of the CBI collected in the field. Another limitation is that CBI is not suitable for use in places where the rock area exceeds 50% [23]. However, in the karst region, particularly in the southwestern part of Guizhou, there are many areas severely affected by rocky desertification. This may necessitate the use of other indices to better represent the impact of fire behavior on the ecosystem.

4. Conclusions

This study evaluated the potential of using Gaofen-1 WFV imagery for fire severity assessment in a karst region in Guizhou Province, Southwest China. A comparative analysis was conducted using Sentinel-2 MSI imagery, based on 19 spectral indices. A minimally modified CBI was used in field fire severity collection and 75 sample sites were gathered. Correlation analysis indicates that most of the indices exhibited significant correlations (p < 0.001) with all of the field-collected CBI dataset, except for NDVI, indicating the possibility of using these indices in fire severity assessment in this region. In Gaofen-1 WFV imagery, BAI, dBAI, and RdMSAVI were identified as the top three indices in their respective categories, whereas in Sentinel-2 MSI data, NBR+, dNDVI, and RdNDVI emerged as the highest-performing indices. With the best-performance indices in each category, different fire severity models were built. The overall accuracy of fire severity assessment for the Gaofen-1 WFV imagery ranged from 40% to 44%, whereas that for the Sentinel-2 MSI imagery ranged from 40% to 48%. The accuracy of the middle and high levels is far higher than that of the low level, which might be attributed to the limitation of the validation samples in the low severity level. The RMSE values for the six models were approximately 0.6, ranging from 0.58 to 0.67. Based on the best performance index among all of the indices for each instrument, the fire severity maps derived from Gaofen-1 WFV and Sentinel-2 MSI imagery exhibited similar spatial distributions; however, Sentinel-2 MSI imagery exhibited a more pronounced mosaic pattern, presumably due to its finer spatial resolution. Compared with UAV imagery, the fire severity maps were also found to be reasonable, especially for distinguishing between high and unburned severity levels. The statistical analysis for every fire severity class showed minimal variations in the proportional distribution between the Gaofen-1 WFV and Sentinel-2 MSI datasets, with differences consistently within 5%. The fire severity maps derived from Gaofen-1 WFV and Sentinel-2 MSI indicate that the middle and high severity levels account for 56.3% and 51.6% of this forest fire, respectively. The results indicated that, after being compared with the field-collected CBI dataset and the Sentinel-2 MSI imagery across a forest fire, the Gaofen-1 WFV imagery can be an effective remote sensing data source for fire severity assessment in this region. Higher resolution Gaofen satellite imagery should be used in the future to further improve the accuracy of fire severity assessments in these areas with high heterogeneity. More forest fires that occurred in different grades of rocky desertification areas, which occupy a considerable portion of Guizhou’s karst region, should be studied to test the effectiveness of this method.

Author Contributions

Conceptualization, Y.L. (Yao Liao) and Y.L. (Yun Liu); methodology, Y.L. (Yao Liao); software, Y.S.; validation, Y.L. (Yao Liao) and Y.L. (Yun Liu); formal analysis, Y.L. (Yao Liao): investigation, Y.L. (Yao Liao), Y.L. (Yun Liu), J.Y., and H.L.; resources, J.F.; data curation, Y.L. (Yao Liao); writing—original draft preparation, Y.L. (Yao Liao); writing—review and editing, Z.Z.; visualization, X.L. and F.H.; supervision, Y.L. (Yao Liao); project administration, J.Y.; funding acquisition, Y.L. (Yao Liao). All authors have read and agreed to the published version of the manuscript.

Data Availability Statement

The data used to support the findings of this study are available from the corresponding author upon request. For readers interested in obtaining these data, please contact the corresponding author via email at [mail to: [email protected]]. Subject to relevant regulations and ethical considerations, we will endeavor to provide necessary access to the data.

Acknowledgments

The authors would like to express their sincere gratitude to the European Space Agency and the financial support provided by the Guizhou Provincial Science and Technology Department. We also thank the editor and anonymous reviewers for their insightful and constructive comments.

Conflicts of Interest

The authors declare no conflicts of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

Abbreviations

The following abbreviations are used in this manuscript:

WFV	Wide Field of View
MSI	Multispectral Imager
RMSE	root mean square error
UAV	unmanned aerial vehicle

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1 Geographical location and slope. (a) The geographical location of the Wanshui town fire events in Guizhou province, China. (b) Gaofen-1 WFV RGB image: red band 2, 0.52–0.59 μm, green band 4, 0.77–0.89 μm, and blue band 3, 0.63–0.69 μm. (c) One real image that was taken from the bottom of the eastern side of the mountain. (d) The slope distribution of the mountain.

Figure 2 Comparison of the spectral response curves of Sentinel-2B MSI and Gaofen-1 WFV sensors.

Figure 3 Field sampling site distribution map.

Figure 4 Photos showing examples of different levels of fire severity among our experimental plots.

Figure 5 Methodological steps.

Figure 6 Scatter plot of the three best indices of Gaofen-1 WFV (a,c,e) and Sentinel MSI (b,d,f) fire severity models based on Gaofen-1 WFV.

Figure 7 Predicted CBI versus field CBI on validation datasets from Gaofen-1 WFV (a,c,e) and Sentinel-2 MSI (b,d,f).

Figure 8 Fire severity maps and fire severity area statistics (e,f) derived from the (a) Gaofen-1 WFV RdMSAVI and (b) Sentinel-2 MSI RdNDVI models, compared with UAV imagery after fire (c,d).

Figure 9 Remapped fire severity map from the Gaofen-1 WFV RdMSAVI model.

Table 1

Summary of the Gaofen-1 and Sentinel-2B datasets used in this study.

Fire Name	Occurrence Date	Remote Sensor		Pre-Fire Date		Post-Fire Date		Field Survey Date
Wanshui town	3 April 2023	Gaofen-1	Sentinel-2B	Gaofen-1	Sentinel-2B	Gaofen-1	Sentinel-2B
		WFV	MSI	3 November 2022	6 November 2022	22 October 2023	22 October 2023	22 August 2023 to 20 October 2023

Table 2

Spectral indices used in fire severity assessments from Gaofen-1 WFV and Sentinel MSI.

Spectral Index	Gaofen-1 WFV	Sentinel MSI	Reference	Categories
Normalized differential vegetation index (NDVI)	${\displaystyle \frac{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}3}{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4+\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}3}}$	${\displaystyle \frac{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4}{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8+\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4}}$	Rouse et al. [50]	Spectral indices
Burned area index (BAI)	${\displaystyle \frac{1}{{(0.1-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}3)}^2+{(0.06-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4)}^2}}$	${\displaystyle \frac{1}{{(0.1-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4)}^2+{(0.06-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8)}^2}}$	Chuvieco et al. [51]
Global environmental monitoring index (GEMI)	${\mathsf{\eta }}_1^{*}(1-0.25)-{\displaystyle \frac{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}3-0.125}{(1-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}3)}}$	${\mathsf{\eta }}_2^{*}(1-0.25)-{\displaystyle \frac{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4-0.125}{(1-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4)}}$	Pinty and Verstraete [52]
Modified soil adjusted vegetation index (MSAVI)	${\displaystyle \frac{2\ast {\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}}_4+1-\sqrt{{(2\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4+1)}^2-8(\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4-\mathrm{b}\mathrm{a}\mathrm{n}\mathrm{d}3)}}{2}}$	${\displaystyle \frac{2\ast {\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}}_8+1-\sqrt{{(2\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8+1)}^2-8(\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8-\mathrm{b}\mathrm{a}\mathrm{n}\mathrm{d}4)}}{2}}$	Qi et al. [53]
Normalized burn ratio (NBR)	—	${\displaystyle \frac{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}12}{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8+\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}12}}$	López-García and Caselles [54]
Normalized burn ratio plus (NBR+)	—	${\displaystyle \frac{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}12-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8\mathrm{a}-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}3-\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}2}{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}12+\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8\mathrm{a}+\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}3+\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}2}}$	Alcaras et al. [55]
Differenced normalized differential vegetation index (dNDVI)	${\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}-{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}$		García-Llamas et al. [42]	Differenced spectral indices
Differenced burned area index (dBAI)	${\mathrm{B}\mathrm{A}\mathrm{I}}_{\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}-{\mathrm{B}\mathrm{A}\mathrm{I}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}$		—
Differenced global environmental monitoring index (dGEMI)	${\mathrm{G}\mathrm{E}\mathrm{M}\mathrm{I}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}-{\mathrm{G}\mathrm{E}\mathrm{M}\mathrm{I}}_{\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}$		—
Differenced modified soil adjusted vegetation index (dMSAVI)	${\mathrm{M}\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}-{\mathrm{M}\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}}_{\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}$		—
Differenced normalized burn ratio (dNBR)	—	${\mathrm{N}\mathrm{B}\mathrm{R}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}$ − ${\mathrm{N}\mathrm{B}\mathrm{R}}_{\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}$	Mallinis et al. [41]
Differenced normalized burn ratio plus (dNBR+)	—	${\mathrm{N}\mathrm{B}\mathrm{R}+}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}-{\mathrm{N}\mathrm{B}\mathrm{R}+}_{\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{t}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}$	—
Relative differenced normalized differential vegetation index (RdNDVI)	${\displaystyle \frac{\mathrm{d}\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}{\sqrt{\left|{\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}\right|}}}$		—	Relatively differenced spectral indices
Relative differenced burned area index (RdBAI)	${\displaystyle \frac{\mathrm{d}\mathrm{B}\mathrm{A}\mathrm{I}}{\sqrt{\left|{\mathrm{B}\mathrm{A}\mathrm{I}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}\right|}}}$		—
Relative differenced global environmental monitoring index (RdGEMI)	${\displaystyle \frac{\mathrm{d}\mathrm{G}\mathrm{E}\mathrm{M}\mathrm{I}}{\sqrt{\left|{\mathrm{G}\mathrm{E}\mathrm{M}\mathrm{I}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}\right|}}}$		—
Relative differenced modified soil adjusted vegetation index (RdMSAVI)	${\displaystyle \frac{\mathrm{d}\mathrm{M}\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}}{\sqrt{\left|{\mathrm{M}\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}\right|}}}$		—
Relative differenced normalized burn ratio (RdNBR)	—	${\displaystyle \frac{\mathrm{d}\mathrm{N}\mathrm{B}\mathrm{R}}{\sqrt{\left|{\mathrm{N}\mathrm{B}\mathrm{R}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}\right|}}}$	Miller and Thode [30]
Relative differenced normalized burn ratio plus (RdNBR+)	${\displaystyle \frac{\mathrm{d}\mathrm{N}\mathrm{B}\mathrm{R}+}{\sqrt{\left|{\mathrm{N}\mathrm{B}\mathrm{R}+}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}\right|}}}$		—
Relativized burn ratio (RBR)	${\displaystyle \frac{\mathrm{d}\mathrm{N}\mathrm{B}\mathrm{R}}{({\mathrm{N}\mathrm{B}\mathrm{R}}_{\mathrm{p}\mathrm{r}\mathrm{e}\mathrm{f}\mathrm{i}\mathrm{r}\mathrm{e}}+1.001)}}$		Parks et al. [20]

${\mathsf{\eta }}_1^{*}={\displaystyle \frac{2({\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4}^2-{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}3}^2)+1.5\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4+0.5\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}3}{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4+\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}3+0.5}}$, ${\mathsf{\eta }}_2^{*}={\displaystyle \frac{2({\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8}^2-{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4}^2)+1.5\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8+0.5\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4}{\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}8+\mathrm{B}\mathrm{a}\mathrm{n}\mathrm{d}4+0.5}}.$

Table 3

Pearson correlation coefficients and significance (p) between the spectral indices of Gaofen-1 WFV, Sentinel MSI, and CBI. Maximum correlation values for each category from each satellite are shown in bold.

Spectral Indices	Gaofen-1 WFV	Sentinel MSI	Categories
NDVI	0.290 *	−0.074	Spectral indices
BAI	−0.515 ***	−0.482 ***
GEMI	0.489 ***	0.457 ***
MSAVI	0.472 ***	0.419 ***
NBR	—	−0.484 ***
NBR+	—	0.525 ***
dNDVI	0.486 ***	0.625 ***	Differenced spectral indices
dBAI	0.583 ***	0.612 ***
dNBR	—	0.445 ***
dNBR+	—	0.444 ***
dMSAVI	0.558 ***	0.461 ***
dGEMI	0.566 ***	0.463 ***
RdNDVI	0.504 ***	0.634 ***	Relatively differenced spectral indices
RdBAI	0.558 ***	0.551 ***
RdNBR	—	0.444 ***
RdNBR+	—	0.470 ***
RdMSAVI	0.593 ***	0.515 ***
RdGEMI	0.579 ***	0.487 ***
RBR	—	0.445 ***

The levels of significance are indicated as * p < 0.05 and *** p < 0.001.

Table 4

Number of samples used for training and validation.

CBI Threshold	[0, 0.25]	(0.25, 1.25)	[1.25, 2.0)	[2.0, 3.0]	Total
Training Points	6	7	26	11	50
Validation Points	3	5	9	8	25

Table 5

Linear models for fire severity assessment for Gaofen-1 WFV and Sentinel-2 MSI using the training dataset.

Model	Model Formula		Correlation Coefficient
Gaofen-1 WFV
RdMSAVI	$\mathrm{C}\mathrm{B}\mathrm{I}=0.0043 \ast \mathrm{R}\mathrm{d}\mathrm{M}\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}+1.2605$	$\mathrm{R}\mathrm{d}\mathrm{M}\mathrm{S}\mathrm{A}\mathrm{V}\mathrm{I}=75.83 \ast \mathrm{C}\mathrm{B}\mathrm{I}-68.709$	$0.572$
dBAI	$\mathrm{C}\mathrm{B}\mathrm{I}=0.0022 \ast \mathrm{d}\mathrm{B}\mathrm{A}\mathrm{I}+1.3874$	$\mathrm{d}\mathrm{B}\mathrm{A}\mathrm{I}=156.14 \ast \mathrm{C}\mathrm{B}\mathrm{I}-202.75$	$0.583$
BAI	$\mathrm{C}\mathrm{B}\mathrm{I}=-0.0161 \ast \mathrm{B}\mathrm{A}\mathrm{I}+2.0825$	$\mathrm{B}\mathrm{A}\mathrm{I}=-18.253 \ast \mathrm{C}\mathrm{B}\mathrm{I}+66.532$	$0.542$
Sentinel-2 MSI
RdNDVI	$\mathrm{C}\mathrm{B}\mathrm{I}=0.0076 \ast \mathrm{R}\mathrm{d}\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}+1.8211$	$\mathrm{R}\mathrm{d}\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=56.003 \ast \mathrm{C}\mathrm{B}\mathrm{I}-131.64$	0.650
dNDVI	$\mathrm{C}\mathrm{B}\mathrm{I}=0.0099 \ast \mathrm{d}\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}+1.8297$	$\mathrm{d}\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}=43.092 \ast \mathrm{C}\mathrm{B}\mathrm{I}-101.79$	$0.653$
NBR+	$\mathrm{C}\mathrm{B}\mathrm{I}=0.0051 \ast \mathrm{N}\mathrm{B}\mathrm{R}++4.3379$	$\mathrm{N}\mathrm{B}\mathrm{R}+=63.659 \ast \mathrm{C}\mathrm{B}\mathrm{I}-663.76$	$0.568$

Table 6

Thresholds of different fire severity classes from models.

Severity ClassModel	Unburned	Low	Middle	High
Gaofen-1 WFV
CBI	[0, 0.25]	(0.25, 1.25)	[1.25, 2.0)	[2.0, 3.0]
dBAI	≤−163.72	(−163.72, −7.58)	[−7.58, 109.53)	≥109.53
BAI	≥61.97	(43.72, 61.97)	(30.03, 43.72]	≤30.03
RdMSAVI	≤−49.75	(−49.75, 26.08)	[26.08, 82.95)	≥82.95
Sentinel-2 MSI
RdNDVI	≤−117.64	(−117.64, −61.64)	[−61.64, −19.63)	≥−19.63
dNDVI	≤−91.02	(−91.02, −47.93)	[−47.93, −15.61)	≥−15.61
NBR+	≤−647.85	(−647.85, −584.19)	[−584.19, −536.44)	≥−536.44

Table 7

Accuracy evaluation of different spectral indices.

Model Name		Fire Severity Class Accuracy
Gaofen-1 WFV	Unburned	Low	Middle	High	Total accuracy	RMSE
dBAI	33.3	20.0	55.6	37.5	40.0	0.61
BAI	33.3	20.0	44.4	50	40.0	0.67
RdMSAVI	33.3	20.0	55.6	50	44.0	0.58
Sentinel-2 MSI
dNDVI	33.3	0.0	55.6	50.0	40.0	0.61
NBR+	33.3	40.0	22.2	75.0	44.0	0.67
RdNDVI	66.7	0.0	66.7	50.0	48.0	0.59

Appendix A

Table A1

Field samples.

Site No.	ABC *	DE *	ABCDE *	Site No.	ABC *	DE *	ABCDE *
1	2.06	/	2.06	39	1.96	/	1.96
2	2.18	2.31	2.21	40	2.17	2.36	2.22
3	2.16	/	2.16	41	1.89	2.26	1.98
4	2.28	/	2.28	42	2.2	/	2.2
5	1.72	/	1.72	43	1.61	1.15	1.32
6	2.04	/	2.04	44	1.84	1.42	1.74
7	1.84	/	1.84	45	2.04	1.8	1.98
8	2	0.88	1.72	46	2.1	2.19	2.12
9	1.92	0.25	1.5	47	2.1	2.2	2.13
10	2.13	/	2.13	48	2.2	/	2.2
11	1.68	0.27	1.33	49	2	/	2
12	1.74	0.46	1.42	50	1.63	0.59	1.28
13	1.81	0.49	1.48	51	1.59	1.01	1.4
14	1.87	0.5	1.41	52	0.45	0	0.34
15	1.61	0.24	1.27	53	0.12	0	0.09
16	1.48	/	1.48	54	0.89	0	0.67
17	2.06	/	2.06	55	0	0	0
18	2.27	/	2.27	56	0	0	0
19	1.81	/	1.81	57	0.15	/	0.15
20	1.89	/	1.89	58	0.06	/	0.06
21	1.33	0.09	0.91	59	0.02	/	0.02
22	2.23	0.72	1.73	60	0.12	/	0.12
23	1.14	0.09	0.88	61	0.45	/	0.45
24	1.72	0.42	1.4	62	0.38	/	0.38
25	1.44	0.43	1.19	63	0.82	/	0.82
26	1.59	0.14	1.11	64	0	/	0
27	1.73	/	1.73	65	2.25	/	2.25
28	1.45	/	1.45	66	2.08	0.88	1.78
29	1.56	0.43	1.28	67	2.13	2.46	2.21
30	1.42	1.44	1.43	68	2.11	/	2.11
31	2.1	2.01	2.08	69	2.06	/	2.06
32	2.07	0.75	1.63	70	1.99	/	1.99
33	1.47	1.19	1.4	71	0.44	/	0.44
34	1.94	1.82	1.9	72	1.11	0.08	0.85
35	1.89	/	1.89	73	1.78	0.2	1.39
36	1.72	1.38	1.64	74	0.14	0.02	0.11
37	1.79	/	1.79	75	1.56	0.08	1.19
38	1.97	1.15	1.77

* ABC, DE, and ABCDE represent the averaged CBI for the understory, overstory, and the entire sample plot, respectively.

References

1. Zhao, J.; Yue, C.; Wang, J.; Hantson, S.; Wang, X.; He, B.; Li, G.; Wang, L.; Zhao, H.; Luyssaert, S. Forest fire size amplifies postfire land surface warming. Nature; 2024; 633, pp. 828-834. [DOI: https://dx.doi.org/10.1038/s41586-024-07918-8] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/39322733]

2. Yilmaz, O.S.; Acar, U.; Sanli, F.B.; Gulgen, F.; Ates, A.M. Mapping burn severity and monitoring CO content in Türkiye’s 2021 Wildfires, using Sentinel-2 and Sentinel-5P satellite data on the GEE platform. Earth Sci. Inform.; 2023; 16, pp. 221-240. [DOI: https://dx.doi.org/10.1007/s12145-023-00933-9]

3. Liu, Z.; Yang, J. Quantifying ecological drivers of ecosystem productivity of the early-successional boreal Larix gmelinii forest. Ecosphere; 2014; 5, pp. 1-16. [DOI: https://dx.doi.org/10.1890/ES13-00372.1]

4. Chebykina, E.Y.; Abakumov, E.V.; Kimeklis, A.K.; Gladkov, G.V.; Andronov, E.E.; Dymov, A.A. Wildfires’ Effect on Soil Properties and Bacterial Biodiversity of Postpyrogenic Histic Podzols (Middle Taiga, Komi Republic). Forests; 2024; 15, 145. [DOI: https://dx.doi.org/10.3390/f15010145]

5. Van Der Werf, G.R.; Randerson, J.T.; Giglio, L.; Van Leeuwen, T.T.; Chen, Y.; Rogers, B.M.; Mu, M.; Van Marle, M.J.; Morton, D.C.; Collatz, G.J. . Global fire emissions estimates during 1997–2016. Earth Syst. Sci. Data; 2017; 9, pp. 697-720. [DOI: https://dx.doi.org/10.5194/essd-9-697-2017]

6. Mathews, L.E.H.; Kinoshita, A. Urban Fire Severity and Vegetation Dynamics in Southern California. Remote. Sens.; 2021; 13, 19. [DOI: https://dx.doi.org/10.3390/rs13010019]

7. Chang, Y.; Zhu, Z.; Bu, R.; Li, Y.; Hu, Y. Environmental controls on the characteristics of mean number of forest fires and mean forest area burned (1987–2007) in China. For. Ecol. Manag.; 2015; 356, pp. 13-21. [DOI: https://dx.doi.org/10.1016/j.foreco.2015.07.012]

8. Fang, K.; Yao, Q.; Guo, Z.; Zheng, B.; Du, J.; Qi, F.; Yan, P.; Li, J.; Ou, T.; Liu, J. . ENSO modulates wildfire activity in China. Nat. Commun.; 2021; 12, 1764. [DOI: https://dx.doi.org/10.1038/s41467-021-21988-6]

9. Masinda, M.M.; Li, F.; Liu, Q.; Sun, L.; Hu, T. Forest Fire Risk Estimation in a Typical Temperate Forest in Northeastern China using the Canadian Forest Fire Weather Index, Case study in Autumn 2019 and 2020. Nat. Hazards; 2022; 111, pp. 1085-1101. [DOI: https://dx.doi.org/10.21203/rs.3.rs-481816/v1]

10. Yi, K.; Zhang, J. Spatial distribution and temporal variability of open fire in China. Int. J. Wildland Fire; 2016; 26, pp. 122-135. [DOI: https://dx.doi.org/10.1071/WF15213]

11. Wu, Z.; He, H.S.; Keane, R.E.; Zhu, Z.; Wang, Y.; Shan, Y. Current and future patterns of forest fire occurrence in China. Int. J. Wildland Fire; 2020; 29, pp. 104-119. [DOI: https://dx.doi.org/10.1071/WF19039]

12. Lian, C.; Xiao, C.; Feng, Z.; Ma, Q. Accelerating decline of wildfires in China in the 21st century. Front. For. Glob. Change; 2024; 6, 1252587. [DOI: https://dx.doi.org/10.3389/ffgc.2023.1252587]

13. Su, J.; Gou, X.; Lambers, J.H.R.; Zhang, D.D.; Zheng, W.; Xie, M.; Manzanedo, R.D. Increasing ENSO variability synchronizes tree growth in subtropical forests. Agric. For. Meteorol.; 2024; 345, 109830. [DOI: https://dx.doi.org/10.1016/J.AGRFORMET.2023.109830]

14. Smith-Ramírez, C.; Castillo-Mandujano, J.; Becerra, P.; Sandoval, N.; Fuentes, R.; Allende, R.; Acua, M.P. Combining remote sensing and field data to assess recovery of the Chilean Mediterranean vegetation after fire: Effect of time elapsed and burn severity. For. Ecol. Manag.; 2022; 503, 119800. [DOI: https://dx.doi.org/10.1016/j.foreco.2021.119800]

15. Hultquist, C.; Chen, G.; Zhao, K. A comparison of Gaussian process regression, random forests and support vector regression for burn severity assessment in diseased forests. Remote Sens. Lett.; 2014; 5, pp. 723-732. [DOI: https://dx.doi.org/10.1080/2150704X.2014.963733]

16. Partheepan, S.; Sanati, F.; Hassan, J. Bushfire Severity Modelling and Future Trend Prediction Across Australia: Integrating Remote Sensing and Machine Learning. arXiv; 2024; [DOI: https://dx.doi.org/10.48550/arXiv.2410.02963]

17. Zikiou, N.; Rushmeier, H.; Capel, M.I.; Kandakji, T.; Rios, N.; Lahdir, M. Remote Sensing and Machine Learning for Accurate Fire Severity Mapping in Northern Algeria. Remote Sens.; 2024; 16, 1517. [DOI: https://dx.doi.org/10.3390/rs16091517]

18. White, J.D.; Ryan, K.C.; Key, C.C.; Running, S.W. Remote Sensing of Forest Fire Severity and Vegetation Recovery. Int. J. Wildland Fire; 1996; 6, pp. 125-136. [DOI: https://dx.doi.org/10.1071/WF9960125]

19. Landmann, T. Characterizing sub-pixel landsat ETM plus fire severity on experimental fires in the Kruger National Park, South Africa. S. Afr. J. Sci.; 2003; 99, pp. 357-360. [DOI: https://dx.doi.org/10.1007/s11948-003-0040-y]

20. Parks, S.A.; Dillon, G.K.; Miller, C. A New Metric for Quantifying Burn Severity: The Relativized Burn Ratio. Remote Sens.; 2014; 6, pp. 1827-1844. [DOI: https://dx.doi.org/10.3390/rs6031827]

21. Zheng, Z.; Wang, J.; Shan, B.; He, Y.; Liao, C.; Gao, Y.; Yang, S. A New Model for Transfer Learning-Based Mapping of Burn Severity. Remote Sens.; 2020; 12, 708. [DOI: https://dx.doi.org/10.3390/rs12040708]

22. Chafer, C.J.; Noonan, M.; Macnaught, E. The post-fire measurement of fire severity and intensity in the Christmas 2001 Sydney wildfires. Int. J. Wildland Fire; 2004; 13, pp. 227-240. [DOI: https://dx.doi.org/10.1071/WF03041]

23. Key, C.H.; Benson, N.C. Landscape Assessment: Ground measure of severity, the Composite Burn Index; and Remote sensing of severity, the Normalized Burn Ratio. FIREMON Fire Eff. Monit. Inventory Syst.; 2006; 164, pp. LA-1-LA-51.

24. Roy, D.P.; Boschetti, L.; Trigg, S.N. Remote sensing of fire severity: Assessing the performance of the normalized burn ratio. IEEE Geosci. Remote Sens. Lett.; 2006; 3, pp. 112-116. [DOI: https://dx.doi.org/10.1109/LGRS.2005.858485]

25. Epting, J.; Verbyla, D.; Sorbel, B. Evaluation of remotely sensed indices for assessing burn severity in interior Alaska using Landsat TM and ETM+. Remote Sens. Environ.; 2005; 96, pp. 328-339. [DOI: https://dx.doi.org/10.1016/j.rse.2005.03.002]

26. Wimberly, M.C.; Reilly, M.J. Assessment of fire severity and species diversity in the southern Appalachians using Landsat TM and ETM+ imagery. Remote Sens. Environ.; 2007; 108, pp. 189-197. [DOI: https://dx.doi.org/10.1016/j.rse.2006.03.019]

27. Escuin, S.; Navarro, R.; Fernández, P. Fire severity assessment by using NBR (Normalized Burn Ratio) and NDVI (Normalized Difference Vegetation Index) derived from LANDSAT TM/ETM images. Int. J. Remote Sens.; 2008; 29, pp. 1053-1073. [DOI: https://dx.doi.org/10.1080/01431160701281072]

28. Cocke, A.E.; Fulé, P.Z.; Crouse, J.E. Comparison of burn severity assessments using Differenced Normalized Burn Ratio and ground data. Int. J. Wildland Fire; 2005; 14, pp. 189-198. [DOI: https://dx.doi.org/10.1071/WF04010]

29. Miller, J.D.; Yool, S.R. Mapping forest post-fire canopy consumption in several overstory types using multi-temporal Landsat TM and ETM data. Remote Sens. Environ.; 2002; 82, pp. 481-496. [DOI: https://dx.doi.org/10.1016/S0034-4257(02)00071-8]

30. Miller, J.D.; Thode, A.E. Quantifying burn severity in a heterogeneous landscape with a relative version of the delta Normalized Burn Ratio (dNBR). Remote Sens. Environ.; 2007; 109, pp. 66-80. [DOI: https://dx.doi.org/10.1016/j.rse.2006.12.006]

31. Gajardo, J.; Yáñez, M.A.; Espinoza, S.E.; Carrasco-Benavides, M.; Ormazábal, Y.; Mena, C.A.; Gómez, P.; Garrido, P.B. Comparison of the Absolute and Relative Difference Spectral Indices to Estimate Burn Severity: The Case of Endangered Nothofagus alessandrii (ruil). Ecol. Restor.; 2022; 40, pp. 191-202. [DOI: https://dx.doi.org/10.3368/er.40.3.191]

32. Mitri, G.H.; Gitas, I.Z. Mapping the severity of fire using object-based classification of IKONOS imagery. Int. J. Wildland Fire; 2008; 17, pp. 431-442. [DOI: https://dx.doi.org/10.1071/WF07103]

33. Manso, O.F.; Manso, A.F.; Pastor, C.Q. Combining spectral mixture analysis and object-based classification for fire severity mapping. For. Syst.; 2009; 18, pp. 296-313. [DOI: https://dx.doi.org/10.5424/fs/2009183-01070]

34. Quintano, C.; Fernandez-Manso, A.; Calvo, L.; Marcos, E.; Valbuena, L. Land surface temperature as potential indicator of burn severity in forest Mediterranean ecosystems. Int. J. Appl. Earth Obs. Geoinf.; 2015; 36, pp. 1-12. [DOI: https://dx.doi.org/10.1016/j.jag.2014.10.015]

35. Zheng, Z.; Zeng, Y.; Li, S.; Huang, W. A new burn severity index based on land surface temperature and enhanced vegetation index. Int. J. Appl. Earth Obs. Geoinf.; 2016; 45, pp. 84-94. [DOI: https://dx.doi.org/10.1016/j.jag.2015.11.002]

36. Guan, D.; Su, W. Quantitative Assessment of Eco-Environment Vulnerability in Karst Region. Lowl. Technol. Int.; 2016; 16, pp. 45-53. [DOI: https://dx.doi.org/10.14247/lti.16.1_45]

37. Li, S.-L.; Liu, C.-Q.; Chen, J.-A.; Wang, S.-J. Karst ecosystem and environment: Characteristics, evolution processes, and sustainable development. Agric. Ecosyst. Environ.; 2021; 306, 107173. [DOI: https://dx.doi.org/10.1016/j.agee.2020.107173]

38. Jiang, Z.; Liu, H.; Wang, H.; Peng, J.; Meersmans, J.; Green, S.; Quine, T.; Wu, X.; Song, Z. Bedrock geochemistry influences vegetation growth by regulating the regolith water holding capacity. Nat. Commun.; 2020; 11, 2392. [DOI: https://dx.doi.org/10.1038/s41467-020-16156-1]

39. Bin, W.; Ming, L.; Dan, J.; Suju, L.; Qiang, C.; Chao, W.; Yang, Z.; Huan, Y.; Jun, Z. A Method of Automatically Extracting Forest Fire Burned Areas Using Gf-1 Remote Sensing Images. Proceedings of the IGARSS 2019—2019 IEEE International Geoscience and Remote Sensing Symposium; Yokohama, Japan, 28 July–2 August 2019.

40. Zhao, Y.; Huang, Y.; Sun, X.; Dong, G.; Li, Y.; Ma, M. Forest Fire Mapping Using Multi-Source Remote Sensing Data: A Case Study in Chongqing. Remote Sens.; 2023; 15, 2323. [DOI: https://dx.doi.org/10.3390/rs15092323]

41. Mallinis, G.; Mitsopoulos, I.; Chrysafi, I. Evaluating and comparing Sentinel 2A and Landsat-8 Operational Land Imager (OLI) spectral indices for estimating fire severity in a Mediterranean pine ecosystem of Greece. GIScience Remote Sens.; 2018; 55, pp. 1-18. [DOI: https://dx.doi.org/10.1080/15481603.2017.1354803]

42. García-Llamas, P.; Suárez-Seoane, S.; Fernández-Guisuraga, J.M.; Fernández-García, V.; Fernández-Manso, A.; Quintano, C.; Taboada, A.; Marcos, E.; Calvo, L. Evaluation and comparison of Landsat 8, Sentinel-2 and Deimos-1 remote sensing indices for assessing burn severity in Mediterranean fire-prone ecosystems. Int. J. Appl. Earth Obs. Geoinf.; 2019; 80, pp. 137-144. [DOI: https://dx.doi.org/10.1016/j.jag.2019.04.006]

43. Saulino, L.; Rita, A.; Migliozzi, A.; Maffei, C.; Allevato, E.; Garonna, A.P.; Saracino, A. Detecting Burn Severity across Mediterranean Forest Types by Coupling Medium-Spatial Resolution Satellite Imagery and Field Data. Remote Sens.; 2020; 12, 741. [DOI: https://dx.doi.org/10.3390/rs12040741]

44. Huangping County. Introduction to Huangping County. Available online: https://www.qdnhp.gov.cn/zjhp/hpjj/ (accessed on 8 August 2024).

45. ESA. Copernicus Data Space Ecosystem. Available online: https://dataspace.copernicus.eu/ (accessed on 10 March 2023).

46. ESA. Overview of Sentinel-2 Mission. Available online: https://sentiwiki.copernicus.eu/web/s2-mission#MSI-Instrument (accessed on 15 March 2024).

47. CNSA. Land Observation Satellite Data Services. Available online: https://data.cresda.cn/ (accessed on 25 May 2023).

48. Santis, A.D.; Chuvieco, E. Burn severity estimation from remotely sensed data: Performance of simulation versus empirical models. Remote Sens. Environ.; 2007; 108, pp. 422-435. [DOI: https://dx.doi.org/10.1016/j.rse.2006.11.022]

49. Zheng, Z.; Zeng, Y.; Li, S.; Huang, W. Mapping Burn Severity of Forest Fires in Small Sample Size Scenarios. Forests; 2018; 9, 608. [DOI: https://dx.doi.org/10.3390/f9100608]

50. Rouse, J.W.; Haas, R.H.; Schell, J.A.; Deering, D.W. Monitoring Vegetation Systems in the Great Plains with Erts. Third Earth Resources Technology Satellite-1 Symposium; NASA: Washington, DC, USA, 1974; Volume 1, pp. 309-317.

51. Chuvieco, E.; Mart, M.P.; Palacios, A. Assessment of different spectral indices in the red-near-infrared spectral domain for burned land discrimination. Int. J. Remote Sens.; 2002; 23, pp. 5103-5110. [DOI: https://dx.doi.org/10.1080/01431160210153129]

52. Pinty, B.; Verstraete, M.M. GEMI: A non-linear index to monitor global vegetation from satellites. Vegetatio; 1992; 101, pp. 15-20. [DOI: https://dx.doi.org/10.1007/BF00031911]

53. Qi, J.; Chehbouni, A.; Huete, A.R.; Kerr, Y.H.; Sorooshian, S. A modified soil adjusted vegetation index. Remote Sens. Environ.; 1994; 48, pp. 119-126. [DOI: https://dx.doi.org/10.1016/0034-4257(94)90134-1]

54. García, M.J.L.; Caselles, V. Mapping burns and natural reforestation using thematic Mapper data. Geocarto Int.; 1991; 6, pp. 31-37. [DOI: https://dx.doi.org/10.1080/10106049109354290]

55. Alcaras, E.; Costantino, D.; Guastaferro, F.; Parente, C.; Pepe, M. Normalized Burn Ratio Plus A New Index for Sentinel-2 Imagery. Remote Sens.; 2022; 14, 1727. [DOI: https://dx.doi.org/10.3390/rs14071727]

56. Soverel, N.O.; Perrakis, D.D.; Coops, N.C. Estimating burn severity from Landsat dNBR and RdNBR indices across western Canada. Remote Sens. Environ.; 2010; 114, pp. 1896-1909. [DOI: https://dx.doi.org/10.1016/j.rse.2010.03.013]

57. Liao, Y.; Li, X.; Huang, L.; Tian, P.; Gu, X. Burned land retrieval from GAOFEN-1 satellite image based on vegetation index. J. Nat. Disasters; 2021; 30, pp. 199-206. [DOI: https://dx.doi.org/10.13577/j.jnd.2021.0519]

58. Tan, L.; Zeng, Y.; Zheng, Z. An adaptability analysis of remote sensing indices in evaluating fire severity. Remote Sens. Iand Resour.; 2016; 28, pp. 84-90. [DOI: https://dx.doi.org/10.6046/gtzyyg.2016.02.14]