An Improved Framework for Forest Fire Severity Assessment in Mountainous Areas Based on the dNBR Index: A Case Study from Central Yunnan, China

Highlights

What are the main findings?

• While the first spring after a fire was identified as the optimal observation window, cross-sensor application of dNBR without recalibration caused a 39% underestimation of high-severity areas, underscoring the critical need for sensor-specific thresholds.
• Topographic correction offered limited practical benefits. Furthermore, dNBR demonstrated reliable applicability primarily in forests with pre-fire NDVI > 0.5.

What are the implications of the main findings?

• This study developed a standardized dNBR evaluation framework that optimizes temporal, sensor, topographic, and vegetation factors, thereby improving the stability and accuracy of dNBR-based forest fire severity classification.
• This framework is also applicable to other remote sensing-based differenced normalized indices, supporting broader utility in change detection studies.

Abstract

Forest fires pose a considerable threat to the security of ecosystems and human society, rendering accurate assessments of fire severity critical for ecological recovery and effective fire management. The differenced Normalized Burn Ratio (dNBR) has been employed to evaluate forest fire severity; however, it presents notable uncertainties owing to variations in data sources, temporal phases, and environmental factors. To address these challenges, this study analyzed 10 forest fires occurring between 2006 and 2023 in central Yunnan Province, China. First, a rapid sampling method utilizing very high-resolution imagery was developed to assess the performance of dNBR classification under varying conditions. Second, the study identified the optimal post-fire observation window and compared classification thresholds and accuracy between Landsat and Sentinel-2 imagery in assessing fire severity. Finally, the research explored the impacts of topographic correction and pre-fire vegetation differences on classification outcomes. The findings revealed the following: (1) Imagery captured in the spring of the fire year, characterized by minimal vegetation interference, demonstrated the highest classification stability and superior capability for identifying high-severity burns. (2) Landsat outperformed Sentinel-2 in regional accuracy (0.92 vs. 0.87), and direct threshold transfer between sensors resulted in a 39% underestimation of high-severity areas, underscoring the necessity for sensor-specific calibration. (3) Topographic correction provided limited practical benefits, merely yielding a marginal improvement in accuracy (+1.44%) with the SCS+C model in steep terrain, and was generally unnecessary. (4) The influence of pre-fire vegetation was discovered to be threshold-dependent: dNBR performed reliably in forests with pre-fire NDVI > 0.5, while adjusted approaches were solely recommended for sparse or heterogeneous vegetation. Overall, this study establishes a systematic framework for optimizing dNBR-based severity assessment, enhancing its accuracy and operational utility in forest fire management.

1. Introduction

Forest fires pose a significant global threat to ecological security, leading to substantial economic losses and disrupting ecosystem structure and functional stability [1]. Alarmingly, the global frequency and intensity of severe wildfires have more than doubled since 2003 [2]. Generally, the ecological impact of forest fires is predominantly influenced by their severity, as more intense fires result in increased tree mortality and higher carbon emissions, thus prolonging ecosystem recovery [3]. Fire severity, defined as the degree of environmental change induced by fire, is a crucial metric for quantifying this short-term ecological impact [4]. Fire severity is a widely used metric for characterizing the impacts of fire on ecosystems. However, confusion persists regarding the distinctions between fire severity, burn severity, and fire intensity. “Fire intensity” refers to the physical combustion process, specifically the energy released from burning organic matter. In contrast, “fire severity” describes the short-term impacts of fire on the environment, whereas “burn severity” refers to the long-term ecological changes introduced to a landscape by fire and the associated ecosystem responses [4]. As a crucial metric for quantifying short-term ecological impact, fire severity is fundamental for evaluating ecological consequences and guiding effective post-fire management. Consequently, the accurate assessment of fire severity is fundamental for evaluating ecological consequences and guiding effective post-fire management.

Remote sensing has become the primary method for evaluating fire severity, largely owing to the remote nature and extensive spatial scale of wildfires [5]. Medium-resolution satellite imagery, particularly data from Landsat, is commonly used for fire severity mapping through various Normalized Burn Ratio (NBR) indices. These indices include the differenced NBR (dNBR) [6], Relative differenced NBR (RdNBR) [7], and Relativized Burn Ratio (RBR) [8]. Specifically, the dNBR index has gained widespread acceptance across diverse ecosystems, such as forests, grasslands, and wetlands, owing to its operational efficiency and reliability [9,10]. Research has confirmed the effectiveness of dNBR for delineating burn scars and assessing wildfire severity [11]. Additionally, it has been integrated into national monitoring initiatives, including the Burn Area Emergency Response (BAER) [12] and Burn Severity Monitoring Trends (MTBS) [11] projects.

However, the dNBR index encounters several challenges that limit its consistency and broader applicability. A significant limitation is the lack of standardized protocols for image selection, threshold determination, and data processing, which considerably undermines the comparability of results across studies [7,13]. Although optimal temporal windows have been suggested [6,9,14], their applicability across diverse geographical regions warrants further validation. Concerning the use of classification thresholds in fire severity mapping, it is widely acknowledged that optimal dNBR thresholds vary spatially owing to geographic heterogeneity [15,16,17]. Nevertheless, the degree to which these thresholds differ across various remote sensing datasets remains unclear. The direct transfer of Landsat-derived dNBR thresholds to other satellite data, such as Sentinel-2, is common [18,19], despite inadequate analysis of inter-sensor spectral and spatial response differences. Additionally, no clear consensus has been reached regarding the necessity of topographic correction or how to account for pre-fire vegetation influences during data processing. While early research indicated that indices such as NBR are relatively insensitive to topography [20], more recent studies contend that terrain and shadowing can significantly bias assessment outcomes, thereby advocating for implementing topographic correction [21,22,23]. In particular, one study specifically examined the impact of topographic correction on the dNBR index and discovered that the modified c-correction could enhance classification accuracy by 0.13 [24]. Similarly, the role of pre-fire vegetation remains a contentious issue. There is general agreement that it complicates dNBR-based assessments [7,8], leading to the development of relative indices such as the RdNBR [7]. Nonetheless, empirical comparisons reveal that the RdNBR does not consistently outperform the dNBR [19,25], questioning the rationale for its widespread adoption. This suggests that the mechanisms underlying vegetation influence are multifaceted and necessitate more nuanced investigation.

These challenges associated with dNBR are further amplified in mountainous environments, where intricate topography exacerbates the inherent limitations of remote sensing techniques. Central Yunnan, China, exemplifies a mountainous area that has become a high-frequency fire zone owing to its unique geographic and climatic conditions. According to official statistics, 1248 forest fires were recorded in this region during the 2005–2014 period [26]. However, the complex topography and limited image availability in this area have led to a considerable shortage of reliable fire severity data. Addressing this gap necessitates the establishment of a mapping method for assessing the severity of forest fires in Central Yunnan.

Therefore, this study aimed to enhance the accuracy and consistency of forest fire severity assessments by examining the key factors influencing the dNBR index and optimizing its application process. The research objectives are to (1) identify the optimal temporal window for post-fire imagery in Central Yunnan; (2) quantify the classification differences between Landsat and Sentinel-2 imagery, as well as determine region-specific optimal dNBR thresholds for fire severity classification; (3) assess the necessity and performance of topographic correction methods in mountainous terrains; (4) explore the impact of pre-fire vegetation conditions on dNBR assessment. By addressing these objectives, this research ultimately seeks to establish a more robust and regionally adapted framework for dNBR-based severity assessment, thereby promoting more reliable fire management and ecological recovery strategies.

2. Materials and Methods

2.1. Study Area

Central Yunnan was selected as the study area, located in southwestern China (101°30′–104°30′E, 24°30′–26°30′N). This region encompasses Kunming, Yuxi, Qujing, and Chuxiong Yi Autonomous Prefecture (Figure 1), covering 27.92% of Yunnan Province’s total area and serving as its demographic and economic core. Typically a plateau region, it has an average elevation of 1800 m. The climate is defined as subtropical monsoon, with distinct wet and dry seasons; over 85% of precipitation occurs from May to October. Winter and spring are marked by drought, posing a high fire risk [27]. Furthermore, the vegetation in this area is diverse, primarily consisting of fire-prone coniferous forests featuring species such as Pinus yunnanensis, P. armandii, and P. kesiya. The dense undergrowth of weeds and shrubs in these forests leads to a significant fuel load, heightening fire sensitivity [28]. Additionally, the implementation of forest conservation projects by the Yunnan government has improved vegetation cover, resulting in fuel accumulation and increasing wildfire risk [29]. Consequently, combining extensive forest coverage and drought-prone conditions creates an environment highly conducive to wildfire ignition and spread, rendering this region a high-risk area for recurrent large-scale fire events [30].

This study analyzed 10 forest fires that occurred in Central Yunnan (Figure 1), 90% of which occurred during the spring drought season (March to May). The areas burned during these events ranged from 51 to 5684 hectares and included diverse vegetation types, such as coniferous, broadleaf, and mixed forests. The perimeters of Fires 2–10 were mapped using very high-resolution (VHR) satellite imagery (spatial resolution > 0.54 m), while the perimeter for Fire 1 was obtained from official government records.

Table 1. Summary of fire characteristics: date, location, VHR resolution, burned area, vegetation types, and sample size.

Fire ID	Fire Date	Location	VHR Image Resolution (m)	Burned Area (ha)	Vegetation Type	Calibration Samples	Validation Points
1	2006-03	Kunming	NA	1849	Broadleaf forest and Coniferous forest	NA	NA
2	2010-02	Kunming	0.14	135	Shrubs and Grassland	48	30
3	2012-03	Yuxi and Kunming	0.54	1592	Shrubs and Coniferous forest	124	56
4	2014-03	Yuxi	0.14	66	Broadleaf forest and Shrubs	55	38
5	2014-04	Kunming	0.27	1263	Broadleaf forest and Shrubs	103	86
6	2014-03	Chuxiong	0.14	96	Broadleaf forest	64	44
7	2014-03	Chuxiong	0.27	51	Broadleaf forest	53	40
8	2014-03	Yuxi	0.27	960	Coniferous forest, Shrubs, and Grassland	101	69
9	2020-05	Kunming	0.27	1017	Coniferous forest and Shrubs	85	66
10	2023-04	Yuxi	0.50	5684	Coniferous forest and Shrubs	92	89

provides detailed information for each fire, including location, burned area, and validation data.

2.2. Data and Pre-Processing

Landsat series imagery (Landsat 5/8/9) served as the primary data source for this study. Additionally, Sentinel-2A images for Fires 9 and 10 were included to facilitate cross-sensor comparisons. To establish scientifically robust temporal baselines, we acquired Landsat scenes spanning a 4-year period (2 years before and 2 years after each fire). In total, 120 images that satisfied strict cloud cover criteria (<20%) were selected. For Sentinel-2A imagery, optimal image pairs were selected after determining the optimal temporal windows. All images underwent rigorous preprocessing, which included radiometric calibration, atmospheric correction using the FLAASH model, and geometric alignment with VHR imagery (root mean square error < 0.5 pixels) to guarantee spatial consistency across the multi-source data. The fire area delineated by the VHR imagery was subsequently used as the spatial boundary for all subsequent index calculations. Finally, based on the VHR imagery, the boundary of each fire area is accurately delineated by visual interpretation, which will be used as the spatial range of all subsequent index calculations.

2.3. Methodology

A detailed workflow was established for this study (Figure 2). First, sample points were constructed using VHR imagery from Google Earth and Jilin-1 Fusion. Second, we determined the optimal post-fire temporal window via multi-temporal analysis. Third, within this optimal temporal window, the dNBR for Landsat and Sentinel-2A was calculated, with the preferred data source and classification thresholds identified through accuracy comparisons. Following this, the effects of different topographic correction methods on the classification results were evaluated. Finally, the spatial correlation between the pre-fire Normalized Difference Vegetation Index (NDVI) and fire severity was analyzed to understand how the pre-fire vegetation background influenced the assessment results.

In total, 1243 sample points were collected and categorized into two sets: 725 calibration samples for threshold segmentation and vegetation background analysis, and 518 validation samples for assessing the accuracy of the classification method. To evaluate the impact of topographic correction on classification accuracy more precisely, a high-density sampling scheme was implemented for Fires 5, 8, and 10, yielding an additional 1000, 1000, and 1970 points, respectively (3970 points in total). This separate set of sample points, designated for validating topographic correction, was designed to ensure an even distribution across various severity levels, aspects, and slope gradients to minimize potential bias.

2.3.1. Sample Point Construction Based on VHR Imagery

To ensure an adequate sample size, sampling points were manually interpreted and collected using VHR imagery from Google Earth Pro and Jilin-1 Fusion. The sampling procedure was as follows: first, initial point data were generated from 30-m resolution Landsat raster data, which were then screened to balance severity class proportions, selecting their image centroids as the final sampling points. This step guaranteed spatial consistency between VHR sample locations and source Landsat data (Figure 3). Second, the severity level of each sample point was ascertained via visual interpretation, based on the color and texture patterns within a 30 m radius. Finally, the assigned severity levels were recorded in the attribute table of the sample points. Additionally, for Fires 9 and 10, the same method was applied to conduct visual interpretation within a 10 m radius at the same sample point locations, with the results intended for accuracy validation of the dNBR utilizing the Sentinel-2A data. The reliability of this visual interpretation was evaluated by assessing inter-interpreter consistency. Three independent interpreters, each with prior remote sensing experience and trained on the classification criteria, independently classified 92 sample points from Fire 10. The resulting overall agreement rate was 89.1%, with a Fleiss’ Kappa value of 0.89, reflecting excellent agreement and confirming the reliability of the sample dataset.

2.3.2. dNBR Calculation and Optimal Post-Fire Image Identification

Previous research has demonstrated that the burned area and severity are primarily influenced by the timing of post-fire imagery, with optimal pre-fire images being acquired during either the spring of the fire year or the preceding spring [31,32]. In light of these findings, this study selected the nearest available spring image as the pre-fire dataset to calculate its NBR index. Subsequently, all usable post-fire images from the fire year and the subsequent year were identified, and their NBR values were calculated. Each optimal pre-fire image was paired with every post-fire image to create a time series of dNBR indices. This approach resulted in a comprehensive dNBR dataset for the 10 fire events across multiple temporal windows. The formulas for calculating NBR and dNBR [6] are as follows.

$$NBR=\left(NIR-SWIR2\right)/\left(NIR+SWIR2\right)$$

(1)

where NIR represents near-infrared reflectance and SWIR2 denotes shortwave infrared reflectance.

$$dNBR={NBR}_{pre}-{NBR}_{post}$$

(2)

where $NB{R}_{pre}$ and $NB{R}_{post}$ signify pre- and post-fire NBR values, respectively.

To further analyze the impact of seasonal variation on the assessment results, the time-series dNBR data were arranged according to the acquisition season of the post-fire imagery. Owing to data availability constraints, the post-fire sequences were categorized into spring, autumn, and winter imagery from the fire year and the subsequent year. However, persistent cloud cover in summer renders optical images unusable, limiting the post-fire sequence to scenes from spring, autumn, and winter of both the fire year and the subsequent year. Before determining the final severity classification thresholds, a preliminary classification was conducted based on empirical thresholds validated for Central Yunnan [31,32]: a dNBR value < 0.2 was classified as unburned, while values $\ge 0.2$ were divided into three fire severity levels (0.2–0.4, 0.4–0.6, and >0.6).

Subsequently, for these seasonal groupings, the optimal window for post-fire image acquisition was identified via a comprehensive evaluation of four key metrics: (1) the accuracy of dNBR-derived burned area against the VHR reference; (2) the coefficient of variation (CV) of the burned area, which quantifies result stability [33]; (3) the mean dNBR value, a direct indicator of fire severity used to assess overall severity levels; (4) the proportional area of high-severity burn, reflecting the ability to accurately identify severely burned areas. These metrics were computed for each seasonal group to allow for comparison.

$$AreaAccuracy=1-\left|A_{\_ref}-A_{\_dNBR}\right|/A_{\_ref}$$

(3)

where $A_{\_ref}$ represents the reference burned area, and $A_{\_dNBR}$ denotes the burned area extracted using the dNBR index.

$$CV={\displaystyle \frac{\sigma }{\mu }}\times 100\%$$

(4)

where $\sigma$ signifies the standard deviation, and $\mu$ indicates the mean.

2.3.3. Determination of dNBR Classification Thresholds for Fire Severity

This study employed an empirical thresholding approach based on sample data to establish the dNBR classification thresholds for fire severity. First, dNBR values calculated from Landsat and Sentinel-2A imagery within their respective optimal temporal windows were extracted for the calibration samples interpreted from VHR imagery. Second, a box plot analysis of the dNBR distributions for each severity class was conducted, and the optimal thresholds for a four-level severity classification were defined based on inter-class separability. Finally, Overall Accuracy ($OA$), User’s Accuracy ($UA$), Producer’s Accuracy ($PA$), and the Kappa coefficient ($\kappa$) were calculated using a confusion matrix to evaluate the performance of classification thresholds for fire severity.

$$PA={\displaystyle \frac{TP}{TP+FN}}$$

(5)

$$UA={\displaystyle \frac{TP}{TP+FP}}$$

(6)

$$OA={\displaystyle \frac{TP+TN}{TP+FN+FP+TN}}$$

(7)

where $TP$, $FP$, $TN$, and $FN$ represent true positive, false positive, true negative, and false negative, respectively.

$$\kappa ={\displaystyle \frac{p_0-p_e}{1-p_e}}$$

(8)

where $p_0$ signifies the count of samples correctly classified, and $p_e$ denotes the expected agreement attributed to chance.

2.3.4. Evaluation of Topographic Correction Efficacy

The Teillet, VECA, C-correction, and SCS+C models are among the most commonly used topographic correction models. These models potentially attenuate topographic effects and enhance surface reflectance accuracy (

Table 2. Calculation formulas for topographic correction models.

Topographic Correction Model	Calculation Formula	References
Teillet Model	${\rho }_m=\rho -a·cosi-b+{\rho }_a$	Teillet P M (1982) [34]
VECA Model	${\rho }_m=\rho ·{\rho }_a/\left(a·cosi+b\right)$	Gao Yongnian (2008) [35]
C Model	${\rho }_m=\rho ·\left({\displaystyle \frac{cosz+c}{cosi+c}}\right)$	Teillet P M (1982) [34]
SCS+C Moedl	${\rho }_m=\rho ·\left(cosz·cosS+c\right)/\left(cosi+c\right)$	Soenen S A (2005) [36]

Note: In the table, ${\rho }_m$ is the radiance after topographic correction; $\rho$ is the radiance before topographic correction; z is the solar zenith angle; i is the solar incidence angle; S is the slope; a and b are the slope and intercept of the linear regression line between radiance and $cosi$, respectively; c is the empirical coefficient, obtained by dividing b and a.

). Therefore, the impact of topographic correction on dNBR-based severity classification was evaluated using three forest fires (Fires 5, 8, and 10). These fires were selected for their sizable burned area (>500 ha), representative topography, and comprehensive data records (including VHR and multispectral imagery). They represent a gradient of terrain complexity. Fire 5 typifies the steepest terrain, with more than 61% of its area on slopes > 25°. Fire 8 represents moderately steep topography, where slopes exceed 25° in over 31% of the area. Fire 10 characterizes relatively gentle terrain, where only 13% of the area comprises slopes steeper than 25°. For each fire, the aforementioned models were used to process the pre- and post-fire images of the optimal pair, and the corresponding corrected dNBR values were computed. The classification accuracy of each corrected and uncorrected dNBR product was evaluated against a reference dataset including 3970 samples interpreted from VHR imagery to identify the most effective topographic correction method for the region.

To analyze the effects of topographic correction, the accuracy of fire severity identification was assessed using the following equation:

$$P=(1-\left|{\displaystyle \frac{a-b}{b}}\right|)\times 100\%$$

(9)

where P denotes the overall accuracy of the classification method, a represents the number of samples classified into various fire severity levels using the dNBR index, and b reflects the number of samples classified through visual interpretation of VHR imagery at the same locations.

2.3.5. Analysis of Pre-Fire Vegetation Influence

This study utilized the pre-fire NDVI to characterize pre-fire vegetation conditions. To examine the potential influence of pre-fire vegetation, the pre-fire NDVI was initially categorized into intervals of 0.1 and overlaid with the fire severity map. The distribution of severity levels within each NDVI interval was subsequently analyzed statistically. Additionally, Pearson’s correlation coefficient (r) and the coefficient of determination ($R^2$) were employed to quantitatively evaluate the linear relationship between pre-fire NDVI values and the corresponding fire severity.

$$NDVI=\left(NIR-RED\right)/\left(NIR+RED\right)$$

(10)

where $NIR$ and $RED$ represent the reflectance of near-infrared and red infrared, respectively.

$$r={\displaystyle \frac{{\sum }_{i=1}^n\left(x_i-\overline{x}\right)\left(y_i-\overline{y}\right)}{\sqrt{{\sum }_{i=1}^n{\left(x_i-\overline{x}\right)}^2}\sqrt{{\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}}}$$

(11)

$$R^2=r^2$$

(12)

where $x_i$ represents the pre-fire NDVI value for the i-th pixel, $y_i$ represents the dNBR value for the same pixel in a high-severity burn area, $\overline{x}$ denotes the mean of pre-fire NDVI values, and $\overline{y}$ denotes the mean of dNBR values.

3. Results

3.1. The Optimal Temporal Window of Forest Fire Severity Assessment

As illustrated in Figure 4a, spring imagery from the fire year exhibited the best performance for total burned area extraction, achieving a mean accuracy of 0.92 with a CV of 0.07 (Figure 4b).

This indicates that data from this period were not only considerably accurate but also stable. In contrast, autumn imagery (from the fire year and the subsequent year) performed significantly worse, with a mean accuracy 28% lower than that of spring imagery. The high CV values (0.33 and 0.53, respectively) further indicate extensive variability. Both the mean dNBR and the proportion of high-severity areas (dNBR > 0.6) directly reflect an image’s ability to determine fire severity. Figure 4c demonstrates that the mean dNBR peaked at 0.42 in the spring immediately following the fire, decreased significantly to 0.22 by autumn, and recovered to 0.27 in the spring of the following year. The proportion of high-severity burn areas displayed an even more pronounced decline (Figure 4d). This proportion peaked in the first post-fire spring and subsequently dropped sharply, with the extent of decrease varying according to the fire event; for instance, it declined from 0.58 to 0.02 in Fire 5 and from 0.30 to 0.13 in Fire 6. In several fires (1, 2, 4, and 8), the proportion approached zero by the winter of the fire year, suggesting that imagery from this period could no longer effectively delineate high-severity areas based on the current threshold. Collectively, these four metrics consistently demonstrate that imagery from the spring immediately following the fire outperformed all other periods in terms of burned area extraction accuracy, result stability, and severity discrimination capability.

Figure 5 presents detailed statistical results for Fires 5 and 6, which exhibited the highest availability of post-fire data among the 10 wildfires studied.

As depicted in Figure 5a,b, the spring imagery from the first post-fire year demonstrated optimal performance across all three metrics: total burned area, high-severity area, and the mean dNBR. For Fire 5, these three metrics displayed a significant decline in all subsequent time phases. In contrast, for Fire 6, the mean dNBR in the post-fire winter even surpassed that of the spring image, reaching 0.41; moreover, the burned area was also greater than the spring level. Nevertheless, the proportion of high-severity burns remained 26.8% lower than that in the spring of the same year. Figure 5c–l provide a qualitative comparison of how different post-fire image dates influenced classification results for Fire 5, utilizing paired VHR imagery and dNBR classification maps. The first post-fire spring (2014113) captured the most extensive area of high-severity burn and demonstrated the closest agreement with the VHR imagery in delineating high-severity and unburned areas. The winter image from the fire year (2015052) performed comparably in identifying the total burned area, but significantly underestimated fire severity. Conversely, the two autumn acquisitions (2014280 and 2015324) exhibited substantial misclassification, erroneously categorizing vast portions of the burned area as unburned, rendering them unsuitable for accurate fire severity assessment. Evidence from quantitative metrics and qualitative spatial comparisons converges to indicate that spring imagery acquired in the year of the fire provides the most reliable basis for severity assessment in Central Yunnan, owing to its optimal balance of accuracy, stability, and discriminative capability.

3.2. Optimal dNBR Threshold for Mountainous Forest Fire Severity Classification

To ascertain the optimal classification thresholds, a dataset comprising 725 calibration samples and 518 validation samples was established based on visual interpretation of VHR imagery, from which the corresponding dNBR values were extracted. First, a box plot analysis of the dNBR distributions for each severity class was performed using the calibration samples (Figure 6). The optimal thresholds for the four-level severity classification were subsequently determined by maximizing the separability between adjacent classes (Figure 6 and

Table 3. dNBR-based fire severity classification thresholds for Landsat and Sentinel-2A datasets.

Date Source	Unburned	Low	Moderate	High
Landsat	<0.18	[0.18–0.38)	[0.38–0.60)	>0.60
Sentinel-2A	<0.18	[0.18–0.35)	[0.35–0.53)	>0.53

). The analysis of these distributions revealed unique separability patterns. Overall, the unburned and high-severity classes displayed favorable separability. In contrast, the low- and moderate-severity classes demonstrated partial overlap in some fires (for example, Fires 5 and 6), thus undermining their discriminability. A comparison between the two datasets revealed that while classification thresholds were considerably consistent for the unburned class, notable disparities existed for moderate- and high-severity classes. In particular, the classification thresholds derived from Sentinel-2A data (0.35 and 0.53 for moderate and high severity, respectively) were consistently lower than those from Landsat data (0.38 and 0.60). This discrepancy was markedly pronounced for high-severity areas, where the Sentinel-2A threshold was 0.07 lower.

The accuracy of dNBR index for classifying fire severity was assessed using 518 independent validation samples, with the relevant confusion matrix detailed in

Table 4. Accuracy assessment results derived from confusion matrices for individual fire events.

Fire		2	4	5	6	8	9	10	9	10
Date Source		Landsat							Sentinel-2A
Producer’s Accuracy (PA)	Unburned	1.00	1.00	0.83	0.90	0.85	0.95	0.94	0.84	1.00
	Low	1.00	0.67	0.96	0.89	0.80	0.93	0.83	0.63	0.76
	Moderate	0.73	0.89	0.91	1.00	0.92	0.92	0.95	0.90	0.86
	High	0.89	0.91	0.88	0.87	0.94	0.85	0.96	1.00	1.00
User’s Accuracy (UA)	Unburned	1.00	0.92	1.00	1.00	1.00	1.00	1.00	0.94	0.81
	Low	0.67	0.86	0.81	0.89	0.84	0.88	0.90	0.71	0.90
	Moderate	0.89	0.73	0.80	0.83	0.71	0.73	0.80	0.64	0.89
	High	0.89	1.00	1.00	0.93	0.94	1.00	1.00	1.00	0.91
Overall Accuracy	(OA)	0.87	0.88	0.89	0.91	0.87	0.91	0.92	0.85	0.89
Kappa	($\kappa$)	0.82	0.84	0.85	0.88	0.83	0.88	0.89	0.80	0.85

. The accuracy analysis across various fires and severity levels identified distinct patterns in $PA$ and $UA$. $PA$ exhibited considerable variation across different fires and severity levels, lacking a consistent trend. Notably, substantial fluctuations were recorded for low- and moderate-severity classes, with $PA$ values ranging from 0.63 to 1.0. In contrast, $UA$ displayed a more stable pattern: unburned and high-severity classes consistently achieved considerably high $UA$, yielding a mean value of 0.96, whereas low- and moderate-severity classes recorded significantly lower $UA$, with mean values of 0.83 and 0.78, respectively. This suggests a greater likelihood of misclassification for pixels categorized as having moderate fire severity. Additionally, marked differences in classification accuracy were noted between the two data sources. For Fires 9 and 10, where both datasets were analyzed, Landsat data exhibited superior overall classification performance, achieving an $OA$ of 0.92 and a $\kappa$ value of 0.89—both surpassing the corresponding values for Sentinel-2A data ($OA$ = 0.87, $\kappa$ = 0.83). This performance disparity was most pronounced among the moderate- and low-severity classes.

3.3. Effects of Topography and Pre-Fire Vegetation on dNBR-Based Fire Severity Assessment

The statistics on severity classification accuracy under various topographic corrections are presented in

Table 5. Accuracy comparison of terrain correction methods (Uncorrected, VECA, Teillet, SCS+C, C) for fire severity classification across fire severity levels.

Fire ID	Severity	Uncorrected	VECA	Teillet	SCS+C	C
Fire #5	Unburned	100%	100%	100%	100%	100%
	Low	75.56%	71.11%	73.33%	77.78%	66.67%
	Moderate	65.91%	70.45%	63.64%	72.73%	63.64%
	High	89.58%	87.50%	88.54%	88.54%	86.46%
	Total	82.69%	81.73%	81.25%	84.13%	78.85%
Fire #8	Unburned	100%	95.83%	93.75%	100%	100%
	Low	73.17%	65.85%	63.41%	65.85%	68.29%
	Moderate	65.38%	55.77%	51.92%	50.00%	50.00%
	High	70.83%	60.42%	60.42%	62.50%	58.33%
	Total	77.25%	70.37%	68.78%	69.31%	68.78%
Fier #10	Unburned	90.31%	81.44%	81.65%	88.87%	77.32%
	Low	82.51%	79.12%	78.44%	79.80%	78.61%
	Moderate	74.87%	70.89%	70.89%	74.65%	77.35%
	High	88.11%	83.08%	83.08%	87.94%	91.12%
	Total	83.95%	78.66%	78.54%	82.81%	81.10%

Note: In the table, bold values represent the highest total classification accuracy for each individual fire.

. For Fire 5, the SCS+C method marginally outperformed the uncorrected data by 1.44%, while the uncorrected data itself achieved a higher accuracy (82.69%) than any of the other three correction methods. In the cases of Fires 8 and 10, the uncorrected data yielded superior accuracy compared to all four correction methods. Specifically, the uncorrected accuracy for Fire 8 was 77.25%, with the corrected results reflecting an average decrease of 7.94%. Likewise, for Fire 10, the uncorrected accuracy was 83.95%, with corrections resulting in an average reduction of 3.67%. Furthermore, this trend was consistent across various fire severity levels, indicating that the impact of topographic correction was independent of fire severity. Therefore, the overall advantage of applying topographic correction in this context was negligible and varied significantly across different fire events.

Figure 7a illustrates the relationship between pre-fire NDVI and dNBR-based fire severity classification. In areas classified as unburned or low-severity, dNBR values remained relatively stable and exhibited minimal dependence on the pre-fire NDVI. Conversely, moderate- and high-severity burn areas displayed a significantly positive correlation between dNBR and pre-fire NDVI, a trend that was particularly evident in high-severity zones. Specifically, as the pre-fire NDVI increased from 0.4 to 0.9, the corresponding median dNBR values rose from 0.41 to 0.89. Notably, in regions where the pre-fire NDVI was below 0.5, the measured dNBR values associated with high-severity burns fell short of the optimal threshold (0.60) established in Section 3.2, suggesting potential limitations in applying this threshold to sparsely vegetated areas. Figure 7b further illustrates the positive correlation between the dNBR and pre-fire NDVI within high-severity zones (r = 0.75, $R^2$ = 0.56), indicating that more favorable pre-fire vegetation conditions are linked to higher dNBR values following high-severity burning.

4. Discussion

4.1. Optimal Temporal Windows for Reliable dNBR Assessment

It is widely acknowledged that image acquisition timing has a critical impact on dNBR values. Nonetheless, recommendations for the optimal temporal window vary across different studies. The developers of the dNBR method have indicated that imagery captured within 2–4 weeks post-fire produces models with optimal explanatory power ($R^2\approx 0.82$). They also noted that model fit and practical classification accuracy can significantly decline beyond 8 weeks owing to vegetation recovery [6]. Based on these findings, Noemi et al. suggest that greater stability in dNBR is achieved when pre-fire images are taken within $\pm 4$ weeks of the same seasonal period in the current or previous year, and post-fire images are captured within 1–6 weeks after the fire [14]. However, these recommendations primarily focus on temporal proximity and provide limited guidance regarding seasonal variation. Chen et al. emphasized that seasonal influences outweigh temporal intervals, advising that both pre- and post-fire images should be selected from the growing season [9]. Similarly, Donato et al. advocated for the use of growing season imagery for pre-fire conditions to enhance the spectral separability between pre- and post-fire states [37].

Contrary to some earlier conclusions, this study indicates that the growing season is not the ideal time frame for assessing fire severity in Central Yunnan, despite confirming the significant impact of seasonal factors. As illustrated in Figure 4 and Figure 5, summer images were rendered completely unusable owing to persistent cloud cover. While autumn images are technically part of the growing season, they produced the least reliable assessments, with area accuracy declining by an average of 30%. Moreover, the proportion of high-severity burns experienced a maximum decrease of 57% (Fire 5). In contrast, spring imagery from the immediate post-fire year yielded substantial benefits: it achieved the highest accuracy in burned area extraction (0.92), exhibited strong stability (CV: 0.07), provided the best capability for extracting high-severity pixels, and had the highest image availability rate (64%) [31]. This trend is strongly associated with the seasonal vegetation phenology of central Yunnan. The rapid post-fire recovery of herbaceous plants and shrubs introduces systematic errors into the dNBR, thereby compromising the reliability of assessments based on imagery from the growing season. Consequently, we recommend avoiding imagery from the growing season when constructing the dNBR for forest fire severity assessment in Central Yunnan and suggest using the first post-fire spring as the optimal time for image selection.

4.2. The Effect of dNBR Calculated from Different Datasets on Forest Fire Severity Assessment

Vegetation indices generated from Landsat and Sentinel-2 imagery exhibit a strong correlation, particularly the NBR, with an $R^2$ value reaching 0.99 [38]. Considering this high correlation, forest fire severity assessments utilizing Sentinel-2 imagery frequently rely on dNBR thresholds established from Landsat imagery [18,19]. Our findings also confirm a robust overall agreement, with an $R^2$ value of 0.91 for the dNBR between the two datasets, indicating a high degree of fit (Figure 8a). Nevertheless, the frequency of Sentinel-2A dNBR values declines more sharply within the high-value range (>0.6), exhibiting a notably lower occurrence in intervals above 0.8 compared to Landsat (Figure 8b). This discrepancy limits the ability of Sentinel-2A data to detect high-severity burned areas when using uncalibrated, universal thresholds. This phenomenon mirrors the pattern reported by Hammill et al. [39]. A comparison of NDVI differences derived from Landsat 7 and SPOT 2 revealed that SPOT 2 identified a lower proportion of high-severity areas than Landsat (10% vs. 25%), despite comparable estimates of total burned area.

Therefore, our study underscores the importance of sensor-specific threshold calibration for accurate dNBR-based severity assessment. The effectiveness of this calibration is illustrated in Figure 8c–j. While the delineation of unburned areas and the total burned area was similar between the uncalibrated and calibrated classifications, their severity categorizations differed significantly. This discrepancy was particularly evident in high-severity burn areas, where the uncalibrated classification underestimated the extent of high-severity burns by approximately 39% compared to the calibrated results. The calibrated Sentinel-2A classification demonstrated greater consistency with Landsat classification and VHR reference imagery. These observed differences potentially originate from variations in sensor characteristics, such as spectral response functions and radiometric resolution. Consequently, we contend that directly transferring dNBR thresholds across sensor platforms without validation and calibration can lead to substantial errors in severity grading. Future applications of the dNBR will expand to include more multispectral and hyperspectral sensors, necessitating the implementation of threshold calibration specific to each data source to enhance the accuracy of severity assessments.

Moreover, the classification accuracy of forest fire severity using Landsat data was slightly higher than that of Sentinel-2A (Table 4), challenging the conventional notion that higher spatial resolution always leads to improved classification performance [40]. Comparable findings have been reported in previous studies; for example, one study indicated that Sentinel-2A achieved inferior accuracy (41.67%) compared to Landsat-8 (47.44%), attributing this discrepancy to the timing of Sentinel imagery acquisition [18,41]. In contrast to previous explanations that linked such differences to acquisition timing, this study controlled for seasonal and phenological effects by utilizing imagery from matched optimal windows. Accordingly, the accuracy advantage of Landsat observed here suggests underlying sensor-derived factors, which we analyze from two perspectives. First, Landsat dNBR values displayed a wider distribution, especially in the high-value range, indicating greater sensitivity to fire-induced spectral alterations, potentially due to differences in spectral bandwidth design between the two sensors. Second, the compatibility between reference data and pixel scale must be considered: the higher spatial resolution of Sentinel-2A imposes stricter requirements on georeferencing accuracy and pixel decomposability. If reference data are not sufficiently adapted to this finer resolution, additional uncertainty may be introduced.

4.3. Response of Forest Fire Severity Assessment to Topographic Correction and Pre-Fire Vegetation

Contrary to established expectations, our results reveal limited benefits of topographic correction in assessing fire severity. Fire 5 exhibited a marginal improvement in accuracy (1.44%) with SCS+C correction; however, uncorrected data consistently demonstrated superior performance in classification accuracy and stability for the other two fire events (Table 5). This counterintuitive finding aligns with the work of Peng et al. [42], who mapped the burned area using the moderate resolution imaging spectroradiometer dataset. In their study, uncorrected data achieved an accuracy of 96.59%, which is comparable to that of the SCS+C (96.69%), while significantly outperforming the Teillet (90.32%) and VECA (91.53%) corrected results. These findings indicate that topographic correction may not universally enhance fire-related remote sensing products. Additionally, we analyzed three potential factors that could explain the discrepancies in the performance of topographic corrections. First, the effectiveness of these corrections appears to be slope-dependent. Veraverbeke et al. [24] established that corrections are most beneficial in steep terrain (slopes > 25°), where accuracy improves as the slope gradient increases. Likewise, Soenen and Vanonckelen [36,43] noted that the SCS+C model can significantly reduce reflectance variability in such areas. Consistent with this slope-dependent pattern, the marginal improvement with SCS+C correction in our study was confined to the steepest terrain: for Fire 5—where over 61% of the area had slopes exceeding 25°—the SCS+C correction method outperformed the uncorrected data and the other three correction methods. Conversely, for Fires 8 and 10, where the proportion of area with slopes exceeding 25° was considerably lower (31% and 13%, respectively), the uncorrected data yielded better results than all four topographic correction methods. Second, the calculation of image ratios (for example, NBR and dNBR) inherently mitigates some topographic effects [20,44,45], potentially diminishing the relative benefits of explicit correction in moderately sloped areas. Finally, latitudinal differences in solar illumination potentially play a crucial role [46]. This latitudinal effect may help explain the contrast between the necessity of topographic correction in mid- to high-latitude regions, where terrain-illumination effects are more pronounced [24], and the limited benefits observed in our lower-latitude study area (approximately 25°N).

To further validate these unexpected findings, we employed VHR imagery of Fire 8 as reference data for visual inspections. Figure 9 illustrates the spatial distribution of VHR imagery and the dNBR for Fire 8 using C, SCS+C, Teillet, and VECA topographic correction methods. All four topographic correction methods consistently classified unburned areas more accurately while underestimating high-severity burn areas. A comparative analysis with post-fire VHR imagery revealed superior spatial agreement for uncorrected classification results, implying that topographic correction potentially introduces additional classification errors. Corrected data consistently underestimated fire severity across three distinct shadow-affected observation areas. These findings validate the notion that topographic corrections are not only highly slope-dependent but may also lead to over-correction, thereby introducing additional uncertainty [47]. Notably, we observed decreased shadow effects in burned areas compared to those in pre-fire conditions (Figure 9a,b), possibly indicating a weakened effect of post-combustion topography, although this hypothesis warrants further verification.

Beyond slope-induced variations, this study revealed a clear aspect-dependent effect in topographic correction. The results demonstrated a strong complementary relationship between sunlit and shaded slopes, leading to a significant reorganization of fire severity distribution. As depicted in Figure 10, this spatial redistribution was particularly pronounced in unburned and high-severity areas. Specifically, all four correction methods increased the proportion of unburned pixels on shaded slopes by 0.23–0.37 while decreasing it on sunlit slopes by 0.08–0.16.

In contrast, high-severity areas exhibited an inverse pattern. Minimal changes (<0.1) were observed on neutral aspects for most methods, except under the C-correction. Notably, this spatial reorganization occurred while the total area of each severity class remained stable (mean change: 0.06), suggesting that a compensatory effect across slope aspects obscured significant internal restructuring. This observed aspect dependence aligns with and extends prior recognition of directional bias in correction methods. For instance, previous research has revealed that the C-correction potentially enhances classification for west-to-north-facing slopes but diminishes accuracy for east-to-south-facing aspects [24]. Therefore, ignoring this effect can introduce considerable bias. Collectively, these findings underscore the necessity for accuracy assessments of topographic correction to shift from a singular focus on changes in total area to a comprehensive analysis of spatial distribution, particularly the differential responses influenced by slope and aspect.

Prior research indicates that pre-fire vegetation conditions, specifically biomass and canopy cover, significantly affect dNBR calculations [7]. Under similar fire severity conditions, forested canopies generally produce higher post-fire dNBR values than shrublands or grasslands [6,48]. This discrepancy potentially diminishes the ability of the dNBR to differentiate between areas with mixed vegetation [7]. The more nuanced findings of this study revealed no significant variation in dNBR values across vegetation types in unburned and low-severity areas, while pronounced differences emerged in moderate- and high-severity regions. In high-severity areas, in particular, the dNBR exhibited a strong linear correlation with the pre-fire NDVI (Pearson’s $r=0.75$), suggesting that better pre-fire vegetation conditions correspond to higher dNBR values. In regions where NDVI was <0.5, such as non-vegetated zones, grasslands, sparse shrubs, and agricultural lands, dNBR values remained consistently lower. This potentially elicited an underestimation of fire severity when applying the proposed threshold rules. Conversely, classification accuracy was unaffected in areas with an NDVI > 0.5. In the fire area studied, the proportion of land with an NDVI < 0.5 was less than 3%; hence, the influence of pre-fire vegetation can be considered negligible. This finding suggests that for forest-dominated fires, pre-fire vegetation can be overlooked, a conclusion supported by Tran [49], who also found the dNBR to attain higher accuracy in resprouting open forests and woodlands. Nonetheless, the application of normalized or relative indices, such as RdNBR, is recommended in areas with mixed vegetation. These indices account for pre-fire vegetation conditions, providing more accurate post-fire assessments while minimizing background interference [7,8].

4.4. Limitations and Perspectives

This study performed a systematic evaluation of the performance of the dNBR based on 10 representative forest fire cases in Central Yunnan. Notwithstanding, several limitations should be acknowledged. First, owing to limitations in accessing field survey data, this study primarily relied on the visual interpretation of VHR imagery to construct the sample dataset. While this method facilitated the efficient acquisition of numerous samples, it inevitably introduced a degree of subjectivity. Second, constrained by the availability of suitable remote sensing imagery, certain analyses were based on a limited number of fire events. For instance, the cross-sensor comparison revealing threshold non-transferability was derived from only two fire events (Fires 9 and 10) using Landsat and Sentinel-2A imagery. This limited sample size may affect the statistical robustness and representativeness of the findings. Moreover, forest environments and fire dynamics display considerable spatiotemporal heterogeneity. The conclusions drawn here are predominantly based on a mid-latitude temperate forest context, and their applicability to diverse ecosystems, such as tropical rainforests and high-altitude regions, has yet to be elucidated. To a certain extent, this limits the spatial extrapolation potential of the findings.

To address these limitations, future research should explore several key directions. First, optimizing sample acquisition methods by integrating field surveys with machine learning techniques could facilitate semi-automatic or automatic sample extraction, thereby enhancing efficiency and objectivity. Second, expanding data sources and case studies to include higher-resolution imagery (for example, WorldView) and a greater number of fire events may strengthen the statistical robustness of conclusions regarding cross-sensor consistency and regional applicability. In particular, the cross-sensor threshold differences observed in this study were derived from a limited number of fire events; therefore, incorporating more diverse fire cases across different regions, vegetation types, and burn severity levels is essential to further validate these findings. Additionally, future work should explore physical correction approaches based on sensor spectral response functions instead of relying solely on empirical calibration, thereby improving the comparability and generalizability of fire severity assessments across multiple sensors. Third, extending the ecological and geographical scope of the research is imperative. Conducting comparative validation across various climatic zones, terrains, and vegetation types may enhance the generalizability and operational applicability of the proposed assessment framework.

5. Conclusions

This study formulated a sampling methodology using VHR imagery to generate fire severity reference data, offering a practical and efficient investigative framework for large-scale forest fire severity assessment. Utilizing this dataset, the accuracy and consistency of dNBR were evaluated across varying temporal windows, data sources, topographic conditions, and pre-fire vegetation statuses. The findings reveal several key conclusions. (1) The optimal temporal window for dNBR-based forest fire severity assessment varies according to region. For Central Yunnan, the optimal post-fire imagery is acquired during the first spring following the fire. Such imagery allows for accurate mapping of burned areas and reliable characterization of fire severity distribution. (2) Landsat data provided superior classification accuracy in this region, yielding a mean overall accuracy of 0.92. Additionally, classification thresholds cannot be directly transferred between different datasets and necessitate independent calibration. (3) The influence of topography demonstrated clear conditional dependency, indicating that the necessity and approach for their correction must be determined based on specific environmental contexts. In low-latitude regions such as Central Yunnan, only the SCS+C method provided marginal improvement in steep terrain, whereas all other correction methods and in other topographic settings, uncorrected data yielded superior performance. These findings highlight the limited necessity of topographic correction in such environments. (4) The dNBR-based fire severity assessment is well suited for areas with dense pre-fire vegetation (NDVI > 0.5). For sparsely vegetated or grassland regions, however, threshold recalibration or the use of relative indices is recommended. This study offers empirical evidence and operational guidance for the standardized and refined application of the dNBR, thereby enhancing its accuracy and reliability in forest fire severity classification. Furthermore, while this study concentrated on systematic evaluation of the dNBR, the results also serve as valuable references for other commonly used diachronic indices, such as RdNBR, dNDVI, and RBR. Considering that these indices share similar calculation methods and application scenarios with the dNBR, the refined methods proposed in this study are equally applicable for optimizing and applying these indices.