Introduction

In many areas of the world, large forest fires have become more frequent in recent years. It is expected that there will be an increase in the number of forest fires and amount of burned areas due to climate change and increases in world population (Amatulli et al. 2013). However, fires are a natural process that affect ecosystems (Whitman et al. 2015) and influence biophysical processes at different temporal and spatial scales, from micro-scale impacts (e.g., on a single plant) to broad landscape patterns and processes (Cochrane and Ryan 2009). Forest areas that do not burn or are exposed to low-intensity fires and do not lose vitality contribute to biodiversity by creating heterogeneous spatial structures (Turner and Romme 1994). To preserve this biodiversity, post fire-management should follow a sustainable ecological approach (Baysal et al. 2016). A good way to accurately characterize burned areas is to determine fire severity, considered as the degree of ecological change in both vegetation and soil caused by fire (Kasischke and Bruhwiler 2002). Fire severity is generally accepted as a measure of fuel consumption and mortality. “Fire severity” and “burn severity” are used to determine the extent of environmental change caused by fire (Keeley 2009). The two terms have different ecological and temporal meanings. Fire severity describes short-term effects, while burn severity describes the long-term effects on ecosystems and vegetation (French et al. 2008). While fire severity is assessed immediately or within the first 30–45 days after a fire event (Key and Benson 2006), burn severity may be estimated at different times and seasons (Key 2006).

Fires often occur in large areas which can result in damage assessments taking considerable effort, money, and time. Therefore, remote sensing methods are frequently used after large fires to determine the extent of damage to ecosystems (Key and Benson 2006). Satellite data such as Sentinel-2, Landsat-8/OLI, and MODIS provide the most widely used free imagery datasets for monitoring (Song et al. 2021). Since remote sensing data provide a quantitative assessment of vegetation and soil-related information, these data are also used for fire severity assessment (Luo and Wu 2022).

Depending on the resolution of satellite sensors, several satellite-based spectral indices can be correlated with data from field measurements. The most commonly used indices to detect fire effects through remote sensing are the Normalized Difference Vegetation Index (NDVI) (Mallinis et al. 2018; Ba et al. 2020), the Normalized Burn Ratio (NBR) (Lutes et al. 2006; Miller and Thode 2007; Fernández-García et al. 2018; Ba et al. 2020), the Enhanced Vegetation Index (EVI) (Wu et al. 2015; Ba et al. 2020), the Soil Adjusted Vegetation Index (SAVI) (Wu et al. 2015), the Global Environment Monitoring Index (GEMI) (Ba et al. 2020), the Normalized Difference Water Index (NDWI) (Beltrán-Marcos et al. 2021), the Normalized Difference Moisture Index (NDMI) (Choubin et al. 2017), the Burned Area Index (BAI) (Chuvieco et al. 2002; Ba et al. 2020), the Burned Area Index Modified–LSWIR (BAIML) (Fornacca et al. 2018), the Burned Area Index Modified-SSWIR (BAIMS) (Fornacca et al. 2018), the Mid Infrared Burn Index (MIRBI) (McCarley et al. 2018), Char Soil Index (CSI) (Pletsch et al. 2019; Ba et al. 2020), Relative differenced Normalized Burn Ratio (RdNBR), and Relativized Burn Ratio (RBR) (Parks et al. 2014). Spectral reflectance values from remote sensing satellite images depend on fire characteristics and vegetation type (French et al. 2008). Severity varies with fire intensity, residence time, tree sizes and species-related physiological characteristics (Michaletz and Johnson 2007; Valor et al. 2017). This results in different spectral responses and therefore different fire severity values that need to be calibrated (Miller and Quayle 2015). To produce fire severity maps from satellite images alone, without the knowledge of site or vegetation conditions, is difficult. Therefore, calibrations using field measurements are necessary. The most widely used method is the Composite Burn Index (CBI) developed by Key and Benson (2006). The CBI divides the forest into five vertical strata and rates them with numerical scores from 0 (unburned) to 3 (completely burned) based on the visual assessment of the amount of fuel consumed, degree of soil scorched, blackening or scorching of trees, and plant regeneration (Key and Benson 2006). In previous studies, the CBI correlated well with spectral reflectance values of remote sensing data (Soverel et al. 2010; Cansler and McKenzie 2012). However, De Santis and Chuvieco (2009) found that the CBI was inconsistent and worked well in some ecosystems but not in others. Therefore, they proposed a modified version, the Geometrically structured Composite Burn Index (GeoCBI). The difference between the CBI and the GeoCBI is that the latter also estimates the fraction of vegetation cover (FCOV). Hence, it is more consistently related to spectral reflectance values than the CBI (De Santis and Chuvieco 2009).

Researchers from different countries have identified spectral thresholds for discriminating fire severity classes in different vegetation types such as steppe (White et al. 1996), tundra (Zhu et al. 2006; Allen and Sorbel 2008), savannah (Alleaume et al. 2005; Borini Alves et al. 2018), meadows (White et al. 1996; Rogan and Franklin 2001), shrubs (White et al. 1996; Epting et al. 2005), temperate forests (French et al. 2008), coniferous forests (Miller and Thode 2007; Mallinis et al. 2018), and deciduous forests (Zhu et al. 2006).

However, only a limited number of studies have investigated the relationship between fire severity data from remote sensing and field measurements of forested areas with dense vegetation cover in the Mediterranean basin, although it is a region heavily affected by forest fire (Lasaponara et al. 2006; Mallinis et al. 2018; Saulino et al. 2020). Further studies are needed that assess fire severity based on regional field measurements correlated with satellite imagery in Mediterranean forest ecosystems (Epting et al. 2005; Hudak et al. 2007; Mallinis et al. 2018).

The purpose of this study was to accurately map fire severity in forest ecosystems of Turkey using local thresholds for the first time. The specific objectives were: (1) to evaluate the performance of different satellite imagery, spectral indices, and field measurements; and (2) to determine classification thresholds of best-performing spectral indices in order to generate fire severity maps to determine the ecological effects of fires in different fire-prone forest ecosystems.

Materials and methods

Study area

This study focused on five forest fires that occurred during 2020 and 2021 (Table 1). Three (Adana-Kozan, Denizli-Çardak, and Antalya-Manavgat) were within the Mediterranean climate region in southern Turkey (Fig. 1), an area at highest risk of forest fires. The other two fires were in Ankara-Nallıhan (western Inner Anatolia) and Kastamonu-Taşköprü (western Black Sea region), both within the western Black Sea climate zone with high fire risk. All fires occurred during the fire season when fire weather conditions prevailed. Fires were selected that had occurred in forests dominated by Turkish pine (Pinus brutia Ten.) and/or Anatolian black pine (Pinus nigra Arn. subsp. pallasiana (Lamb.) Holmboe var. pallasiana), as they are the two main forest types in Turkey most affected by fires. These forests are occasionally mixed with broad-leaved species and maquis shrubland species. The sites were heterogeneous in terms of topography and vegetation.

Table 1 Vegetation and climate characteristics of the study areas
Full size table
Fig. 1
figure 1

Location of the burned areas

Full size image

Imagery and preprocessing

Sentinel-2, Landsat-8/OLI, and MODIS satellite images are freely available and used in this study. Cloudless pre- and post-fire images closest to the fire dates were selected. Imageries were downloaded from USG Earth Explorer server. The Sentinel satellite consists of 13 bands with a revisiting time of 10 days. Landsat-8/OLI is 11 bands with a revisiting time of 16 days while MODIS is 7 bands with a revisiting time of 8 days (Table 2).

Table 2 Spectral characteristics used for estimating fire severity (Fernández-Manso et al. 2016; Choubin et al. 2017; Korhonen et al. 2017)
Full size table

Preprocessing of satellite data included geometric (Itten and Meyer 1993), radiometric (Teillet 1986), and atmospheric (Kaufman and Sendra 1988) corrections. Data were first defined in the same coordinate system using the Universal Transverse Mercator projection and WGS84 datum. Geometric corrections included the elimination of distortions between satellite data using ground control points. Among the geometrically corrected data, the root-mean-square error (RMSE) should be lower than 0.5 (Lunetta and Elvidge 1999). Landsat-8/OLI Collection 2 Level 1 satellite data were downloaded using the Digital Elevation model with radiometric and geometric corrections already applied. Geometric, radiometric, and atmospheric corrections of Sentinel-2 satellite data were made using the Sentinel-2 toolbox (Gascon and Ramoino 2017).

After the data were geometrically corrected, a RMSE of 0.44 was calculated. Radiometric corrections included the elimination of atmospheric effects that might cause data errors and the radiometric calibration of pixels that do not represent reflectance values. Atmospheric corrections used the extrapolated Rayleigh-corrected reflectance at NIR and SWIR bands to derive their ratios and visible aerosol single scattering contributions (aerosol epsilon). FLAASH (Fast Line-of-sight Atmospheric Analysis of Hypercubes) and flat field were used for atmospheric corrections to remove effects such as aerosols and water vapor from satellite data (Matthew et al. 2002; Ye et al. 2016; Ilori et al. 2019). After these corrections were completed, spectral indices from the literature (Table 3) were calculated using pre- and post-fire satellite images and the Raster Calculator tab within the ArcMap software. Fire severity maps were generated from the differences between the two images (Mallinis et al. 2018; Saulino et al. 2020) (Table S1).

Table 3 Spectral indices used for estimating fire severity
Full size table

Field sampling

Fire severity measurements were conducted within 5–45 days following the fire. Sampling plots were randomly selected within the burned area using the stratified random sampling approach. It was ensured that the number of plots was proportional to the area each fire class covered. A total of 478 sampling plots were determined. Depending on the burn size, the number of plots for the Adana-Kozan fire was 70, 60 for Ankara-Nallıhan, 105 for Kastamonu-Taşköprü, 15 for Denizli-Çardak, and 228 for the Antalya-Manavgat fire. At each 30 × 30 m sampling plot, fire severity conditions were visually assessed according to the CBI field protocol proposed by Key and Benson (2006) for all five strata by assigning each a value from 0 (unburned) to 3 (severely burned) (Table 4). At the center of each plot, digital images covering different angles were acquired and GPS coordinates were recorded. The GeoCBI for each plot was assessed following the protocol described by De Santis and Chuvieco (2009).

Table 4 Measured CBI and GeoCBI values of different fire severity classes on some of the sampling plots
Full size table

Relationship between remote sensing and field-measured fire severity

To determine the relationships between field-measured fire severity and the spectral indices from satellite imageries, Kendall’s correlation analysis was performed. Based on these results, subsequent analyses were only performed for GeoCBI and spectral indices from Sentinel-2 and Landsat-8/OLI satellite imageries. Linear regression analysis was used to determine the relationship between GeoCBI and the spectral indices with the highest Kendall’s correlation coefficient (Escuin et al. 2008; Mallinis et al. 2018; García-Llamas et al. 2019).

Defining threshold values for spectral indices

Threshold values were determined for best performing spectral indices. For this, optimal binning analysis was carried out using SPSS for spectral indices with the highest R2 derived from Sentinel-2 and Landsat-8/OLI satellite images. In order to classify the values of the spectral indices, the classification proposed by Miller et al. (2009) was used: unburned (0–0.1), low (0.1–1.25), moderate (1.26–2.25), and high (2.26–3.0). This approach allowed for the identification of reflectance threshold values of the most suitable spectral indices based on the locally measured GeoCBI classification. Kappa statistics were then used to assess agreement between GeoCBI classes and optimized classified threshold values. In addition, for each of these spectral indices, a comparison was made between the predicted classes and those obtained via field measurements, and prediction accuracy percentages were calculated (Tables S2 and S3).

Results and discussion

According to our results, spectral indices from both Sentinel-2 and Landsat-8/OLI satellite imageries correlated best with the GeoCBI measurements. After De Santis and Chuvieco (2009) suggested the use of GeoCBI, noting that CBI does not work well in some regions and/or vegetation types, some studies either used GeoCBI instead of CBI or both, to test which one correlates better with spectral reflectance values. Consistent with our results, Mallinis et al. (2018) found in their study from Greece, that GeoCBI was better correlated with spectral indices from both Landsat-8/OLI and Sentinel-2 satellite imagery than CBI. In a study from Australia (Parker et al. 2015), the dNBR spectral index from Landsat satellite images showed a strong correlation with GeoCBI. However, Cansler and McKenzie (2012) in the USA showed that the use of CBI led to better results than GeoCBI. Saulino et al. (2020), on the other hand, reported that spectral indices from Landsat-8/OLI correlated better with CBI while spectral indices from Sentinel-2 were better correlated with GeoCBI.

Of the 478 fire severity sampling plots from five different fire areas, 78 were unburned, 62 were lightly burnt, 152 were moderately burnt, and 186 showed high fire severity. In general, GeoCBI scores were lower than CBI scores for these sampling sites (Fig. 2). In cases where the sampling plots were unburned or all strata showed similar fire severity, CBI and GeoCBI were the same or similar to each other. However, in cases where the ground cover showed high fire severity but the upper crown layer remained unburned, GeoCBI resulted in lower values than CBI (Table 4). This is because GeoCBI is calculated by weighing each layer (stratum) according to its estimated coverage within the plot. Further, if the upper canopy is dense, it causes most of the satellite observed reflectance (De Santis and Chuvieco 2009). Considering only the upper or lower stratum during fire severity classification can lead to incorrect results. It may therefore be concluded that, in cases where the upper canopy is dense and the level of damage is low, GeoCBI provides a better characterization of fire severity.

Fig. 2
figure 2

Comparison of CBI and GeoCBI values obtained for the respective sampling plots

Full size image

Of the three different satellite data used in this study, Sentinel-2 and Landsat-8/OLI were more closely correlated with field measurement data than MODIS (Table 5). Further, although Landsat-8/OLI has a lower resolution (30 m) than Sentinel-2 (10 m), it provided better results for some of the indices due to its higher resolution, but since the differences between results are usually minor, both Sentinel-2 and Landsat-8/OLI can be used (Mallinis et al. 2018; Saulino et al. 2020). Correlation coefficients between MODIS and field measurements were significantly lower for most spectral indices (Table 5), possibly because of the low resolution of MODIS satellite images (250 m).

Table 5 Kendall’s correlation coefficients between spectral indices and field measurements (CBI and GeoCBI)
Full size table

Linear regression was performed to determine the relationship between the spectral indices that resulted in the highest correlation coefficients (Table 5) and GeoCBI values (Figs. 3 and 4). The ones with the highest coefficients of determination (R2) were dSAVI (R2 = 0.73), and RdNBR (R2 = 0.73) for the Sentinel-2 data as well as dNDVI (R2 = 0.74) and dSAVI (R2 = 0.74) for the Landsat-8/OLI data.

Fig. 3
figure 3

Linear regression between spectral indices (dSAVI, RdNBR, dNBR, dNDVI, dNDMI and dNDWI) from Sentinel-2 satellite images and GeoCBI

Full size image
Fig. 4
figure 4

Linear regression between spectral indices (dSAVI, dNDVI, RdNBR dNBR, and dNDWI) from Landsat-8/OLI satellite images and GeoCBI

Full size image

Based on the GeoCBI classification, fire severity thresholds of the best-performing spectral indices were calculated using optimal binning analysis. Optimal classification was performed for five spectral indices (except dNDWI) derived from Sentinel-2 (Table 6) and for three spectral indices (dSAVI, dNDVI, and dNBR) from Landsat-8/OLI (Table 7).

Table 6 Threshold values of the spectral indices from Sentinel-2 satellite images
Full size table
Table 7 Threshold values of spectral indices from Landsat-8/OLI satellite images
Full size table

According to the literature, most fire severity studies determine thresholds of Landsat dNBR. For example, regional dNBR classification thresholds were determined for Alaska (Epting et al. 2005), Montana, USA (Key and Benson 2006), California, USA (Miller and Thode 2007), Indonesia (Hoscilo et al. 2013), Australia (Parker et al. 2015), and Greece (Mallinis et al. 2018). The Landsat dNBR threshold values from our study are closest to those of Miller and Thode (2007) for California. Our results differ from those of Mallinis et al. (2018) for Greece, although their study was conducted in a Pinus brutia stand on Thasos Island which is close to Turkey with similar vegetation. Compared to other studies, our threshold classification values were overall lower. Regional differences can occur based on fire intensity, vegetation cover and species-specific reaction to the fire, leading to different reflectance values corresponding to the fire severity classes (Miller and Quayle 2015). Thresholds of spectral indices can perform well in terms of fire severity classification for a specific place and time but may not perform well elsewhere (Huang et al. 2016).

Based on Kappa statistics, the level of agreement between spectral indices and field GeoCBI was highest for Sentinel-2—based RdNBR (Kappa values for RdNBR: 0.653, dNBR: 0.622, dSAVI: 0.512, dNDMI: 0.506, and dNDVI: 0.001) and Landsat-8/OLI—based dNBR (Kappa values for dNBR: 0.603, dSAVI: 0.588, and dNDVI: 0.588).

Accuracy percentages were calculated between observed and predicted GeoCBI. For all fire severity classes, highest accuracy was for Sentinel-2—based RdNBR (72.1%) and Landsat-8/OLI—based dNBR (69.2%). RdNBR and dNBR showed highest accuracies for unburned, low, and high severity classes but lower accuracies for the moderate class (Tables S3 and S4). Some studies found that dNBR was less sensitive to consumption in the lower strata while the upper canopy was still green (Hoy et al. 2008; De Santis and Chuvieco 2009) but our results show that RdNBR and dNBR are also successful in the low severity class.

According to the results of this study, spectral indices showing the highest accuracy and agreement with field measured GeoCBI were Sentinel-2—based RdNBR and Landsat-8/OLI—based dNBR. In most of studies that tested the performance of spectral indices to estimate fire severity, NBR—based dNBR and RdNBR proved to be the best in detecting fire-caused changes (Epting et al. 2005; Escuin et al. 2008; Veraverbeke et al. 2011; Cansler and McKenzie 2012; Fernández-García et al. 2018; Ariza et al. 2019; García-Llamas et al. 2019). Other studies showed that RBR provides a more accurate estimation of fire severity than dNBR and RdNBR and thus the use of the spectral index RBR is suggested (Parks et al. 2014; Ariza et al. 2019). However, in our study, RBR performed poorly. The performance of NDVI, which was less than dNBR and RdNBR in our study, varied from superior (Chen et al. 2011) to good (García-Llamas et al. 2019) to poor (Fernández-García et al. 2018) in other studies. Hudak et al. (2007) found that the performance of dNDVI and dNBR were comparable and that both could be used for fire severity classification. In another study by Arnett et al. (2015) for a low-intensity prescribed fire in Canada, dSAVI was the best performing index among others calculated from both Landsat and Rapideye satellite images. They stated that dSAVI better represented areas of low-intensity fires and low fire severity. However, in this study, although dSAVI was among the indices showing the highest correlation with ground measurements for both Sentinel-2 and Landsat-8/OLI, overall agreement of dSAVI and field GeoCBI was inferior and its performance to accurately estimate fire severity classes was poor. dMIRBI is recommended for use in areas with low fire severity as it shows higher spectral separability, especially when a dense canopy is present (McCarley et al. 2018). Based on our results, we were not able to support this conclusion.

NDVI and SAVI are indices using NIR and RED bands, while NBR and NDMI use NIR and SWIR bands. With the reflectance values of the Landsat-8/OLI satellite images, it can be seen that spectral bands 4 (RED), 5 (NIR), 6 (SSWIR), and 7 (LSWIR) show high separability between fire severity classes (Fig. 5). With regards to the reflectance values of Sentinel-2, the bands that reached high separability were bands 4 (RED), 5 (RED Edge1), 6 (RED Edge2), 7 (RED Edge3), 8 (NIR), 8a (NIRn), 11 (SSWIR), and 12 (LSWIR) (Fig. 6). This agrees with results from other studies (Lewis et al. 2007; Papageorgiou et al. 2012; Pleniou and Koutsias 2013; Liu et al. 2016; Ariza et al. 2019; Chuvieco et al. 2019; Luo and Wu 2022) and our study as indices using these bands provided the best results. Overall, in our study, NBR-based RdNBR and dNBR indices were more successful in classifying fire severity than the other indices. This is because the NIR band is sensitive to chlorophyll levels, while SWIR is sensitive to plant and soil water contents as well as to the ash content after a fire (Miller and Thode 2007). However, although dNDVI and dSAVI were highly correlated with field measurements, they did not perform as well as RdNBR and dNBR in terms of fire severity classification. This is because RED bands, although sensitive to fire-related decreases in chlorophyll content, show only limited sensitivity to post-fire attributes such as black carbon or ash (García-Llamas et al. 2019). Further, the better performance of Sentinel-2 based RdNBR over dNBR may be explained by the differences in pre-fire chlorophyll levels and density of vegetation cover. Therefore, RdNBR allows for a more accurate estimate of fire severity in heterogeneous terrains (Miller and Thode 2007).

Fig. 5
figure 5

Reflectance values of different fire severity classes according to Landsat-8/OLI satellite images (B indicates spectral bands)

Full size image
Fig. 6
figure 6

Reflectance values of different fire severity classes according to Sentinel-2 satellite images (B indicates spectral bands)

Full size image

Fire severity maps were generated using the threshold values calculated for the Sentinel-2—based RdNBR (Fig. 7) and Landsat—based dNBR indices (Fig. 8). For all fires, the majority of the burned areas were classified as high fire severity (Tables S4 and S5). This was expected for fires that occurred during the fire season and during periods of extreme fire weather conditions.

Fig. 7
figure 7

Fire severity maps according to the Sentinel-2 based RdNBR thresholds

Full size image
Fig. 8
figure 8

Fire severity maps according to the Landsat-8 based dNBR thresholds

Full size image

Conclusions

Fire severity maps are used to describe fire effects on flora and fauna habitats, soil, water systems, atmosphere, and society. They assist forest managers in post-fire decision-making processes. This study evaluated the performance of field-based estimates (CBI and GeoCBI) and spectral indices derived from Sentinel-2, Landsat-8/OLI, and MODIS images to assess fire severity in heterogeneous forest ecosystems of Turkey. Classification thresholds were calculated for well-performing indices and used for fire severity mapping. According to the results, regardless of the satellite used, NBR-based RdNBR and dNBR indices estimated fire severity more accurately than indices based on RED and NIR bands. While Sentinel-2 and Landsat-8/OLI images produced similar good results, MODIS's performed poorly. GeoCBI was closely related to spectral reflectance values than CBI and thus provided more reliable field measurements of fire severity. Spectral reflectance values can vary substantially depending on fire intensity, vegetation cover and species-specific reaction to the fire, and thus can cause regional differences when determining fire severity. This is why threshold values for dNBR in this study differ from those of other studies. Our results can be used in forest ecosystems with vegetation similar to that in this study and might also guide further research. Further, in order to better understand the effects of fire, especially in patchy, heterogeneous forest ecosystems, further research is needed from different regions and fire types, testing and validating fire severity maps based on remote sensing data.