International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

Contents lists available at ScienceDirect

International Journal of Applied Earth

Observation and Geoinformation
journal homepage: [Link]/locate/jag

Immediate assessment of forest fire using a novel vegetation index and

machine learning based on multi-platform, high temporal resolution remote
sensing images
Hanqiu Xu a,b,* , Jiahui Chen a,b, Guojin He c , Zhongli Lin d, Yafen Bai a,b, Mengjie Ren a,b,
Hao Zhang e , Huimin Yin a,b, Fenfen Liu a,b
a
College of Environment and Safety Engineering, Institute of Remote Sensing Information Engineering, Fuzhou University, Fuzhou 350116, China
b
Fujian Provincial Key Laboratory of Remote Sensing of Soil Erosion and Disaster Prevention, Fuzhou University, Fuzhou 350116, China
c
Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China
d
College of Architecture and Urban Planning, Fujian University of Technology, Fuzhou 350118, China
e
Sichuan Academy of Forestry, Chengdu, Sichuan 610081, China

A R T I C L E I N F O A B S T R A C T

Keywords: Forest fires pose a significant threat to ecosystems, biodiversity, and human settlements, necessitating accurate
Burned area mapping and timely detection of burned areas for post-fire management. This study focused on the immediate assessment
SEVI of a recent major forest fire that occurred on March 15, 2024, in southwestern China. We comprehensively
Random Forest
utilized high temporal resolution MODIS and Black Marble nighttime light images to monitor the fire’s devel­
Ecological impact
Yajiang forest fire
opment and introduced a novel method for detecting burned forest areas using a new Shadow-Enhanced
Vegetation Index (SEVI) coupling with a machine learning technique. The SEVI effectively enhances the vege­
tation index (VI) values on shaded slopes and hence reduces the VI disparity between shaded and sunlit areas,
which is critical for accurately extracting fire scars in such terrain. While SEVI primarily identifies burned forest
areas, the Random Forest (RF) technique detects all burned areas, including both forested and non-forested
regions. Consequently, the total burned area of the Yajiang forest fire was estimated at 23,588 ha, with the
burned forest area covering 19,266 ha. The combination of SEVI and RF algorithms provided a comprehensive
and efficient tool for identifying burned areas. Additionally, our study employed the Remote Sensing-based
Ecological Index (RSEI) to assess the ecological impact of the fire on the region, uncovering an immediate 15
% decline in regional ecological conditions following the fire. The usage of RSEI has the potential to quantita­
tively understand ecological responses to the fire. The findings achieved in this study underscore the significance
of precise fire-burned area extraction techniques for enhancing forest fire management and ecosystem recovery
strategies, while also highlighting the broader ecological implications of such events.

1. Introduction Timely assessment of forest fires is crucial in comprehending the

impact of fires on the natural environment, ecosystems, and human
Forests play an irreplaceable role in maintaining ecological balance, society. In particular, understanding the extent and spatial distribution
safeguarding water and soil resources, and regulating climate. Serving of burned areas immediately after a fire is essential for identifying areas
as reservoirs of biodiversity and carbon sinks, forests provide the that require management to mitigate environmental impacts and facil­
ecosystem services crucial for human survival (Wingfield et al., 2015). itate timber recovery (Lee and Chow, 2015; Stow et al., 2007; Weber
However, forests have faced increasing threats in recent years, with et al., 2008). Traditional methods for assessing forest fires primarily rely
increasingly frequent forest fires globally. These fires can significantly on ground surveys, which demand substantial human and material re­
alter ecosystems by affecting species composition, soil characteristics, sources but offer limited coverage and are susceptible to constraints such
and nutrient cycles. Additionally, they impose significant negative im­ as weather and terrain conditions. In contrast, remote sensing technol­
pacts on human societies and economies (Rashkovetsky et al., 2021). ogy holds significant advantages in forest fire assessment. Satellite

* Corresponding author at: College of Environment and Safety Engineering, Fuzhou University, Fuzhou 350116, China.
E-mail address: hxu@[Link] (H. Xu).

[Link]
Received 4 July 2024; Received in revised form 12 September 2024; Accepted 6 October 2024
Available online 11 October 2024
1569-8432/© 2024 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license ([Link]
nc-nd/4.0/).
H. Xu et al. International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

remote sensing can acquire images of various resolutions over large performance is generally comparable to TCT and supervised classifiers,
areas, enabling rapid assessment of fire situations. Utilizing remote or slightly inferior (Epting et al., 2005; Henry, 2008; Rogan and Yool,
sensing technology facilitates the analysis of fire development trends 2001).
and spread, as well as the accurate identification of burned areas. This In recent years, deep learning (DL) and machine learning (ML) have
provides crucial support for forest fire monitoring, emergency response, been involved in fire monitoring and fire severity assessment (Badhan
and post-fire recovery. Current remote sensing methods mainly rely on et al., 2024; Belgiu and Drăguţ, 2016; Knopp et al., 2020; Matthew et al.,
spectral change information of forest locations before and after fires. 2023). Gibson et al. (2020) mapped fire severity in southeastern
Various remote sensing techniques have been recognized as effective Australia by random forest (RF) classification. Their results achieved
methods for monitoring forest fires and identifying fire-burned areas various accuracies ranging from 70 % to 95 % across different classes.
(Chu and Guo, 2014; Miller et al., 2023). Hultquist et al. (2014) demonstrated that RF outperformed the support
Forest fires can result in a reduction of chlorophyll and canopy vector regression (SVR) technique in burn severity mapping. This was
moisture and exposure of bare soil beneath the canopy, leading to a also confirmed by the work of Lee et al. (2023). Hu et al. (2021) used
decrease in reflectance in the near-infrared (NIR) wavelength region and both DL models (U-Net, HRNet, and Fast-SCNN) and ML (KNN and RF)
an increase in reflectance in the mid-infrared (MIR) wavelength region to map burned areas using Sentinel-2 and Landsat-8 data. They indicated
(Jensen, 2000; van Wagtendonk et al., 2004; Chen et al., 2020). that DL models outperformed ML models in two of three study cases.
Therefore, spectral indices derived from NIR and MIR bands are the However, ML models were found to be better suited for mapping
main techniques used in mapping forest fires (Bretreger et al., 2024; de discrete burns in boreal forests. Rashkovetsky et al. (2021) developed a
Carvalho et al., 2015; Giddey et al., 2022; Miller et al., 2023; Weber U-Net-based workflow for detecting burned areas from Sentinel-1/2/3
et al., 2008). and Moderate Resolution Imaging Spectroradiometer (MODIS) images.
Key and Benson (2001) conducted an assessment of two wildfires in The results indicated that the combination of Sentinel-2 and − 3 images
Montana using Landsat Thematic Mapper (TM) imagery. They con­ provided the highest accuracy (Kappa = 0.76) in clear weather condi­
structed a Normalized Burn Ratio (NBR) derived from NIR and MIR tions, while the combination of Sentinel-1 and − 2 images achieved the
bands and contrasted it with the NIR-based Normalized Difference highest accuracy (Kappa = 0.59) in cloudy weather. Smith et al. (2021)
Vegetation Index (NDVI). Their study revealed that, compared to NDVI, employed both RF and spectral indices (NBR and NDVI) to assess the
NBR exhibited a stronger correlation with field-investigated burn burn severity of two forest fires in Alaska. They revealed that RF out­
severity. Aiming to remove water noise, Alcaras et al. (2022) developed performed the indices at one location (with an overall accuracy of 83 %
an NBR+index by introducing the blue and green bands into the original compared to 73 %) but significantly underperformed the indices at the
NBR formula. The result showed that the index could exclude cloud/ other (67 % vs. 96 %).
water pixels that were erroneously identified as burned areas by other The above analyses indicate that current fire assessments primarily
methods. To assess the spatiotemporal transferability of differenced NBR rely on the methods derived from remote sensing spectral indices/
(dNBR), Soverel et al. (2011) compiled burn severity data from ten transforms and DL/ML models. The use of TCT/PCA-based methods for
wildfires in western Canada. They found that the dNBR model showed forest fire assessment is presently less common. Their accuracy is
transferability across diverse landscapes in western Canada. The burn generally lower than that of NBR-based techniques. Vegetation indices
severity study by Rogan and Yool (2001) found that the Soil-Adjusted such as NDVI and SAVI, which rely mainly on the NIR band, are less
Vegetation Index (SAVI) outperformed NDVI but was less effective effective than NBR, especially in steep mountainous regions, due to their
than the MIR/NIR index and the wetness component of the Tasseled Cap lack of the MIR band and susceptibility to topographic shadows.
Transform. Among the 13 remote sensing indicators examined by Epting Regarding the types of remote sensing imagery used, researchers
et al. (2005), NBR achieved the highest accuracy, whereas the accuracy usually employ only 1–2 image types. There is limited application of
of vegetation indices (VI), such as NDVI and SAVI, were generally coupling multiple-platform remote sensing images. Among the
ranked below the sixth place. Mallinis et al. (2018) compared the per­ employed remote sensing images, Landsat and Sentinel-2 images are
formances of spectral indices (NDVI and NBR) between Landsat-8 favored by users due to their wide wavelength range simultaneously
Operational Land Imager (OLI) and Sentinel 2A for mapping burn covering MIR and NIR regions and having well-developed TCT co­
severity in Kavala, Greece. They found that NBR outperformed NDVI for efficients (Baig et al., 2014; Crist, 1985; Shi and Xu, 2019). On the other
both the Sentinel-2 and Landsat-8 data. These comparative studies have hand, images with only visible-infrared bands are used less frequently,
shown that the accuracy of fire severity mapping based on VIs like NDVI, and the use of nighttime light images is also limited. Additionally, pre­
EVI, and SAVI was generally unsatisfactory, with particularly poor re­ vious studies have seldom evaluated the impact of forest fires on
sults in steep mountainous regions (Gibson, 2020). Lyndersen et al. regional ecological quality.
(2016) noted that areas with dense vegetation cover and rugged terrain, Therefore, this paper aims to use a recent major forest fire in Yajiang
especially those affected by topographic shadows, pose substantial ob­ County, China, as a case study to achieve the following objectives: (1)
stacles to remote sensing techniques aimed at detecting changes, such as Develop a novel technique to improve the accuracy of NDVI-like vege­
assessing burn severity. tation indices for assessing forest fires in steep terrain. (2) Explore the
The Tasseled Cap Transform (TCT) has also been a commonly used use of multi-platform remote sensing images for forest fire monitoring
method to map forest fires as the coefficients of brightness, greenness, and assessment, particularly the potential of images with only visible-
and wetness components in the TCT can capture characteristics of infrared bands. (3) Assess the impact of the forest fire on regional
burned areas (Miller et al., 2023; Stow et al. 2007). The Disturbance ecological conditions. This research will be significant for monitoring
Index (DI) widely used for mapping forest fire disturbances is derived and assessing forest fires, restoring post-fire ecological conditions, and
from the three TCT components (Masek et al., 2008). Patterson and Yool formulating post-fire management strategies.
(1998) utilized all three TCT components for classifying burn severity in
southeastern Arizona, which yielded a Kappa coefficient of 0.72. The 2. Methods
comparison results of Epting et al. (2005) revealed that the TCT
greenness component was often ranked in the top 5 among the study 2.1. Overview of Yajiang forest fire
cases. Stow et al. (2007) employed Landsat TCT to assess the burn
severity of wildfires in southern California. The results based on the bi- Yajiang County, situated in the mountainous plateau region, is
temporal TCT had accuracies of 64 % and 55 % for the two study areas, located in the Garze Tibetan Autonomous Prefecture of Sichuan Prov­
respectively. Another mathematical transformation, Principal Compo­ ince (29◦ 03′ to 30◦ 30′N, 100◦ 19′ to 101◦ 26′E), China. Most areas within
nent Analysis (PCA), has also been employed in wildfire monitoring. Its the county lie at an altitude of over 3000 m, with snow commonly found

2
H. Xu et al. International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

on mountaintops. The average annual temperature is 11.1 ◦ C, and the after the fire. CBERS-04A 2 m panchromatic and 8 m multispectral high-
total annual precipitation is 807.3 mm. resolution images were used for accuracy validation.
Yajiang County, an important forest resource area in Sichuan Prov­ The MODIS imagery used includes its daily atmospherically cor­
ince, has experienced several forest fires in recent years. The most recent rected surface reflectance L2G products (MOD09GA, MOD09GQ,
forest fire occurred in Gaila Town, north of the county seat (Fig. 1). The MYD09GA, and MYD09GQ) and daily land surface temperature L3
fire was initiated at 17:00 on March 15, 2024, and subsequently products (MOD11A1 and MYD11A1). The VNP46A2 product of the
expanded due to extreme gusty winds, leading to the development of a Black Marble Daily Level 3 data was employed, which is the daily lunar
major forest fire ([Link] BRDF (bidirectional reflectance distribution function)-adjusted NTL
9642357&wfr=spider&for=pc). During the main period of the fire, product (Román et al., 2018). Both MODIS and Black Marble products
there was no precipitation in the area. The temperature ranged from 3 to were downloaded from the NASA website ([Link]
25 ◦ C throughout day and night. The prevailing winds came primarily [Link]). Landsat 8/9 Collection 2 Level 2 Surface Reflectance
from the west and east, with an average speed of 5.9 m/s, but often (SR) products and Surface Temperature (ST) products were used, which
reaching 15 to 20 m/s. The forests at the fire sites are predominantly were downloaded from the United States Geological Survey (USGS)
composed of broadleaf evergreen forests including Cyclobalanopsis website ([Link] GF-1 and GF-6 images were
aquifolioides, Quercus semecarpifolia, and Rhododendron parvifolium, and obtained as Level 1A products and downloaded from the China Re­
coniferous forests such as Pinus densata, Abies fabri, and Picea asperata. sources Satellite Application Center website ([Link]
After efforts by over 3000 personnel, the forest fire was completely cn/#/home).
extinguished by 10:00 AM on March 28. This fire affected a total of 837 The downloaded MODIS and Landsat images are already in reflec­
households and 3396 individuals across 11 villages. ([Link] tance or land surface temperature products and can thus be directly
[Link]/s?id=1795215081917895779&wfr=spider&for=pc). used. However, the GF-1 and GF-6 images require preprocessing. The
brightness values (DN) of these images were corrected for atmospheric
effects using official calibration data and the Illumination and Atmo­
2.2. Remote sensing data
spheric Correction Model (IACM) (Xu, 2015). Table 1 outlines the data
used and their respective applications.
To gain a thorough understanding of the fire development, this study
collected various sources of remote sensing imagery available before
and after the event, including MODIS, Black Marble nighttime lights 2.3. Mapping burned areas
(NTL) product, Landsat-8/9, GaoFen-1 (GF-1), GaoFen-6 (GF-6), and
China-Brazil Earth Resources Satellite 04A (CBERS-04A) data. Among 2.3.1. Spectral index-based burned area mapping
these, MODIS daily reflectance products and Black Marble NTL products Among the remote sensing data used in this study, the coarse spatial
provide high temporal resolution images to fully understand the resolution of MODIS imagery is unsuitable for accurately mapping
development of the forest fire. The 16 m resolution images from GF-1 burned areas. Landsat provides images only up to March 23, leaving a 5-
and GF-6 were used to map burned areas and estimate burn severity. day gap before the end of the fire on March 28, which does not accu­
Landsat 8/9 imagery was used to detect ecological changes before and rately represent the post-fire situation. Therefore, we opted for the GF-6

Fig. 1. GF-6 natural-color image of Yajiang County acquired on March 29, 2024, along with photographs depicting the fire scene.

3
H. Xu et al. International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

Table 1
Images used for the study.
NDVI = (NIR − Red)/(NIR + Red) (1)

Sensor Date Local Spatial Application SAVI = [(1 + L)(NIR − Red)]/(NIR + R + L), (L = 0.5) (2)
overpass resolution (m)
time
EVI = 2.5(NIR − Red)/(NIR − 6Red − 7.5Blue + 1) (3)
MODIS Terra 03-06- 10:30 250, 500, Pre-fire
2024 1000
Aqua 03-16- 13:30 250, 500, Post-fire
kNDVI = tanh{[(NIR − Red)/2σ]2 } (4)
2024 1000
Terra 03-17- 10:30 250, 500, Post-fire where Blue, Red, and NIR are the reflectance values of the blue, red, and
2024 1000 NIR bands, respectively. The parameter σ represents a length scale and
Terra 03-18- 10:30 250, 500, Post-fire can be approximated effectively by its average value, σ = 0.5
2024 1000
(NIR+Red). This results in a simplified index expressed as (Camps-Valls
Terra 03-20- 10:30 250, 500, Post-fire
2024 1000 et al., 2021):
Terra 03-23- 10:30 250, 500, Post-fire
2024 1000 kNDVI = tanh(NDVI2 ) (5)
Terra 03-25- 10:30 250, 500, Post-fire
Differenced VIs (dVI) can be utilized to determine vegetation
2024 1000
Terra 03-28- 10:30 250, 500, Post-fire changes before and after fires and thus detect burned areas:
2024 1000
Black Marble NTL 03-15- 2:21 500 Pre-fire
dVI = postVI − preVI (6)
2024
03-16- 3:43 500 Post-fire where preVI and postVI denote the VI values before and after fires,
2024 respectively.
03-17- 3:24 500 Post-fire Many comparative studies have indicated that the accuracy of fire
2024
severity mapping based on VIs such as NDVI, EVI, and SAVI was
03-18- 3:05 500 Post-fire
2024 generally unsatisfactory (Epting et al., 2005; Mallinis et al., 2018), with
03-22- 3:30 500 Post-fire particularly poor results in steep mountainous regions, influenced by
2024 topographic shadows. Our study also found that shadows on shaded
03-23- 3:11 500 Post-fire
slopes resulted in significantly lower values of NDVI-like VIs compared
2024
03-24- 3:18 500 Post-fire
to those on sunny slopes. When performing differencing between pre-
2024 and post-fire images, lower VI values on shaded slopes in the pre-fire
03-29- 3:18 500 Post-fire image lead to less significant differences in burned areas when sub­
2024 tracted from the corresponding post-fire image, causing missed
Landsat- OLI, 01-11- 11:45 30, 100 Pre-fire
detections.
8 TIRS 2024
Landsat- OLI2, 03-23- 11:45 30, 100 Post-fire Recognizing this challenge, we developed a novel Shadow-Enhanced
9 TIRS2 2024 Vegetation Index (SEVI), designed to effectively enhance VI values on
GF-1 WFV 03-11- 11:55 16 Pre-fire shaded slopes. Its equation is:
2024
GF-6 WFV 03-29- 11:58 16 Post-fire SEVI = δ(NIR − Red)/(NIR + Red) = δ*NDVI (7)
2024
CBERS- WPM 03-25- 12:13 2, 8 Validation where δ is a shadow enhancement factor designed to enhance NDVI
04A 2024
values specifically on shaded slopes and thus balance the NDVI-value
difference between sunlit and shaded slopes. δ can be determined
image from March 29 as the post-fire image (one day after the fire ended based on the relationship between VI values observed on sunlit and
burning) and the GF-1 image from March 11 as the pre-fire image. This shaded slopes.
selection enables the utilization of the same GF family images to cover The relationship between NDVI values on sunny and shady slopes
the entire period before and after the fire for mapping burned areas. was established through regression analysis of samples. Each pair of
Additionally, the narrow 18-day interval between the selected pre- and NDVI samples was specifically selected from vegetation on both sunny
post-fire images helps mitigate vegetation phenological errors. and corresponding shady slopes using a region-based sampling method.
Remote sensing spectral indices are most commonly used for map­ Regression analysis was performed using the mean values of these
ping burned areas as they have demonstrated strong correlations with sampled regions to derive equations that describe the relationship. The
fire severity (Mallinis et al., 2018; Miller et al., 2023). The indices are values from shaded slopes were treated as the independent variable,
typically developed using NIR and MIR bands (Epting et al., 2005), and while those from sunny slopes served as the dependent variable. It
the spectral information changes in these bands before and after fires should be noted that during sampling, the vegetation types on shaded
can effectively reveal burned regions (Chu and Guo, 2014; Morresi et al., and sunny slopes should be similar to ensure consistency and avoid
2022). However, the GF-1 and GF-6 images selected for mapping burned scenarios where one side comprises grassland while the other is forested.
areas in this study only include visible and NIR bands and lack the MIR The coefficients obtained from the regression equation can be used as
band. Consequently, algorithms relying on the MIR band, such as the adjustment parameters to calibrate the NDVI values on shady slopes. In
widely used NBR, cannot be applied. Therefore, this study employed this study, the regressions yielded SEVIs for pre-fire GF-1 and post-fire
vegetation indices (VIs) derived from the visible and NIR bands of the GF-6 images as:
GF-1 and GF-6 images for mapping burned areas. The review by Miller SEVIpre− = 0.1182NDVI + 0.8941(r = 0.719, p < 0.001) (8)
fire
et al. (2023) indicates that NDVI (Rouse et al., 1973), SAVI (Huete,
1988), and Enhanced Vegetation Index (EVI, Huete et al., 2002) are the SEVIpost− fire = 0.6242NDVI + 0.3015(r = 0.632, p < 0.005) (9)
three most frequently used VIs in mapping burn severity and were hence
selected in this study. Additionally, the recently proposed kernel NDVI The objective of SEVI is to enhance NDVI values exclusively on shady
(kNDVI, Camps-Valls et al., 2021) was also used to broaden the slopes, while the values on sunny slopes do not require enhancement.
analytical scope. The equations of these four VIs are: Therefore, the successful use of SEVI lies in distinguishing between

4
H. Xu et al. International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

sunny and shady slopes. Since reflectance in the vegetation-related red 2.3.2. RF-based burned area mapping
and NIR bands often significantly decreases on shady slopes, the dif­ Random Forest (RF) is a robust, non-parametric ensemble machine
ferentiation between sunny and shady slopes can be achieved by learning classifier built from decision trees that do not assume specific
examining the numerical distribution of reflectance in these two bands. data distributions (Belgiu and Drăguţ, 2016). It uses bagging to create
Using a box plot can help identify a separation point (threshold) multiple sub-datasets through random sampling and trains different
between sunny and shady slopes. Fig. 2 illustrates the distribution of decision trees on each to enhance model diversity and mitigate over­
reflectance data for the red and NIR bands on sunny and shady slopes fitting (Gislason et al., 2006). RF introduces randomness by selecting a
taking the March 29 GF-6 image as an example. We selected a threshold subset of features at each split, ensuring that each tree differs, which
of 0.12 from the NIR band for differentiation because the minimum of further boosts diversity. Predictions are made by aggregating the results
the NIR band on the sunny slope is 0.121, whereas the maximum on the of all decision trees, either by voting or averaging, which leads to
shaded slope is 0.118, with no overlap between the two slope types reduced overfitting and improved generalization (Collins et al., 2020).
(Fig. 2, right). Due to the overlapping data distributions between sunny RF has been demonstrated to be effective for fire mapping over the
and shady slopes, the red band was excluded from the threshold selec­ past decade (Gibson et al., 2020; Hu et al., 2021; Hultquist et al., 2014).
tion. In cases where there is overlap in either the red or NIR bands, the Therefore, we also employed the RF algorithm to map burned areas from
threshold can be chosen from a position between the upper quartile of the March 29 post-fire GF-6 image. Based on the land cover character­
shaded slope data and the lower quartile of sunny slope data, based on istics of the study area, six categories were initially classified: burned
boxplot separation. area, forest, grass, bare soil, water, and snow. Training samples for the
To prevent misclassification of non-vegetative land cover types with six categories were collected using the class-balanced method (Mellor
NIR band reflectance values below 0.12 as shady slope vegetation, we et al., 2015). A total of 14,537 training samples were collected for the six
also employ a specific NDVI threshold for separating vegetation and categories, meeting the requirement of 0.25 % coverage of the study
non-vegetation. This ensures that the enhancement specifically targets area (Belgiu and Drăguţ, 2016). Subsequently, the six categories were
the vegetation on shady slopes accurately. aggregated into two categories: burned area and non-burned area.
Although the coarse resolution of MODIS imagery makes it difficult We set the number of trees to 500 as the value is mostly recom­
to estimate the burned area accurately, we still used it to estimate the mended, and errors tend to stabilize before reaching this number
daily burned area during the fire to gain a general understanding of the (Gibson et al., 2020; Gislason et al., 2006; Lawrence et al., 2006). The
fire’s progression. The burned area from MODIS imagery was extracted number of variables at each node within a tree was set to the square root
using the NBR index: of the total number of input variables (Belgiu and Drăguţ, 2016; Collins
et al., 2020).
NBR = (NIR − MIR)/(NIR + MIR) (10)

where NIR and MIR are the reflectance of the MODIS NIR (841–876 nm) 2.3.3. Accuracy assessment
and MIR (2105–2155 nm) bands, respectively. The CBERS-04A’s 2 m panchromatic and 8 m multispectral images
The dNBR is used to extract the burned area. The equation for dNBR were fused to generate a 2 m color reference image. Since there are only
is: two categories (burned area and non-burned area) in the classification
results, 300 samples were randomly selected for validation. The accu­
dNBR = preNBR − postNBR (11) racy was assessed using metrics such as overall accuracy (OA), Kappa
coefficient, omission error (OE), and commission error (CE), which are
where preNBR is the pre-fire image from March 6, and postNBR refers to commonly used in assessing the performance of burned area mapping
the MODIS images from the subsequent dates. (Kurbanov et al., 2022).

Fig. 2. Distribution of reflectance data for the red and NIR bands from the March 29 GF-6 image, illustrating the separation of the data between sunlit and shaded
slopes. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

5
H. Xu et al. International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

2.4. Assessing the fire’s impact on eco-environments moisture component of TCT (Zhai et al., 2022), dryness by the
Normalized Difference Built-up and Soil Index (NDBSI) (Xu et al., 2019),
The Remote Sensing Ecological Index (RSEI) (Xu et al., 2018) is and hotness by the Landsat-9 C2L2 Surface Temperature (ST) product.
widely utilized in assessing ecological conditions across diverse These four indicators are combined using the first principal component
geographical environments (Firozjaei et al., 2021; Hu and Xu, 2019; (PC1) to form RSEI, expressed as:
Maity et al., 2022; Qureshi et al., 2020; Xu et al., 2024). In a recent
RSEI = PC1[f(NDVI, Wetness, NDBSI, ST)] (12)
review focusing on assessing ecological vulnerability, Kamran and
Yamamoto (2023) emphasize RSEI’s impact and acceptance within the The detailed equations for calculating the above-mentioned in­
scientific community. They indicate that RSEI is an important concep­ dicators can be found in Xu et al. (2024). RSEI values range from 0 to 1,
tual model and has emerged as an effective ecological assessment model. with higher values signifying better ecological conditions.
This study incorporates RSEI to evaluate the regional ecological
impact of the fire. RSEI integrates four indicators—greenness, wetness, 3. Results
dryness, and hotness—into a comprehensive ecological monitoring and
assessment index. These four indicators can capture the ecological 3.1. Fire spread process
response to the fire. Each of the indicators is represented by a remote
sensing indicator. Greenness is represented by NDVI, wetness by the The precise recognition of the fire development process is essential

Fig. 3. A series of MODIS and Black Marble nighttime light images showing the fire process. MODIS images are presented using pseudocolor composites (RGB: MIR2,
NIR, and green bands) and are further enhanced with the thermal infrared band to highlight the fire. (For interpretation of the references to color in this figure
legend, the reader is referred to the web version of this article.)

6
H. Xu et al. International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

for gaining a comprehensive understanding of how the extent and fre­ 18, it can be seen that the fire spots decreased in number and intensity,
quency of fires influence eco-environments (Lentile et al., 2006). Fig. 3 and their spatial extent was no longer expanding, indicating that the fire
presents a series of MODIS images from March 6 to 28, with the image of had been brought under control. In the subsequent daily images, only
March 6 depicting pre-fire conditions. These images were enhanced sporadic fire spots were observed. The lights in the early morning image
using the thermal infrared band because it can detect and highlight on March 29 returned to levels similar to those before the fire on March
actively burning fires. The thermal emissive power from fires is much 15, indicating that the fire event had ended. The entire sequence of
more intense than that from the surrounding background, creating high nighttime lights during the forest fire process aligns with the MODIS
contrast that enables reliable detection of active fires, even if they cover images.
a tiny fraction of a pixel (e.g., <0.01 %) (Lentile et al., 2006). This
capability is particularly useful for coarse-resolution MODIS thermal 3.2. Extraction of burned areas
infrared images, as it means that the image can detect a hotspot with an
area as small as 1 ha within its 1 km2 pixel. 3.2.1. Enhancement of VI values on shaded slopes
Despite the fire breaking out on the afternoon of March 15, there Fig. 4 presents a comparative evaluation of SEVI versus the other
were no apparent signs of the fire in the image from 1:30 PM on March four VIs used across both sunlit and shaded terrain. The shadows from
16. However, in the image from 10:30 AM on March 17, thick smoke was Fig. 4a were first extracted (Fig. 4b and 4c), and the vegetation infor­
visible on site, with extensive areas showing signs of burning and mul­ mation within these shadows was subsequently enhanced to generate
tiple active fire spots. Nevertheless, the burned area could not be esti­ SEVI (Fig. 4d). SEVI exhibits a reduced disparity between sunny and
mated due to the heavy smoke. Subsequent images from March 18 shaded regions in the image, attributed to its notable enhancement of
onwards showed a significant reduction in the presence of smoke and vegetation details within shaded areas. In contrast, distinctions between
active fire spots, with the burned area not notably expanding beyond sunny and shaded slopes are more pronounced in the remaining four VIs,
that of March 17. The burned area derived using the dNBR approach for with northwest-facing slopes demonstrating significantly lower index
March 18, 20, 23, 25, and 28 was 20,287 ha, 20,317 ha, 20,943 ha, values compared to their sunlit counterparts. Particularly, this differ­
21,100 ha, and 23,401 ha, respectively. It is evident that the peak of this entiation is more pronounced in SAVI and EVI (Fig. 4g and h), while
major forest fire occurred mainly between the afternoon of March 16 NDVI and kNDVI (Fig. 4e and f) also show discernible variations within
and the morning of March 18, after which the fire was brought almost deeply shaded regions.
under control. To quantitatively measure the contrast in the values of the used five
The series of Black Marble NTL images shows that on the early VIs between shaded and sunny slopes, a metric using the ratio of VI
morning image of March 15 (before the fire), there were only two values between shaded and sunny slopes was employed. Through the
relatively distinct lights on the left side of the image. On the early measurement of 5613 sample points collected on both sunlit and shaded
morning image of March 16, sporadic fire spots began to appear, indi­ slopes from eleven sites, the differences in each VI between the shady
cating the onset of the fire but suggesting the situation was not severe. In and sunny slopes were examined. It is found that EVI exhibits the most
the early morning image of March 17, many pixels were missing (null), substantial deviation, with shaded slope values representing only 40.6 %
which may have been caused by overexposure and brightness saturation of those on sunlit slope values, whereas SEVI achieves an average ratio
due to high-glare areas generated by the intense fire flames (cf. Hao value of 99.1 %, reflecting its closer approximation to sunlit slope values
et al., 2023). However, despite the numerous null values, many bright (Table 2). The mean ratio values of the indices rank as follows: SEVI
and widely distributed bright (fire) spots could still be observed, forming (99.1 %) > NDVI (92.7 %) > kNDVI (88.8 %) > SAVI (43.2 %) > EVI
a large outline of the fire area and indicating an exacerbated fire sce­ (40.6 %). SEVI benefits primarily from introducing the shadow
nario and rapid flame propagation. In the early morning image of March enhancement factor, δ. This adjustment offsets the disparity in VI values

Fig. 4. Enhancement effect of SEVI on shaded slopes and its comparison with four other vegetation indices. (a) Original pre-fire GF-1 image dated March 11 (RGB:
NIR, Red, and Green band composite), (b) Highlight of extracted vegetation overlaid on shaded slopes, noting some dark-grey areas representing bare soil that are
therefore not extracted, (c) Extracted vegetation in shaded areas, (d) SEVI, (e) NDVI, (f) kNDVI, (g) SAVI, and (h) EVI. (For interpretation of the references to color in
this figure legend, the reader is referred to the web version of this article.)

7
H. Xu et al. International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

Table 2
Mean values and ratios of the used vegetation indices for shaded and sunny slopes.
Site Aspect and Ratio No. of samples NDVI kNDVI SAVI EVI SEVI

1 Sunlit slope (SLS) 413 0.790 0.553 0.346 0.336 0.790

​ Shaded slope (SDS) 207 0.780 0.542 0.135 0.122 0.816
​ Ratio of SDS/SLS (%) ​ 98.7 98.0 39.0 36.3 103.3

2 Sunlit slope 375 0.825 0.592 0.378 0.370 0.825

​ Shaded slope 113 0.731 0.489 0.167 0.153 0.798
​ Ratio of SDS/SLS (%) ​ 88.6 82.6 44.2 41.4 96.7

3 Sunlit slope 179 0.837 0.605 0.385 0.377 0.837

​ Shaded slope 217 0.746 0.505 0.178 0.164 0.809
​ Ratio of SDS/SLS (%) ​ 89.1 83.5 46.2 43.5 96.7

4 Sunlit slope 453 0.814 0.580 0.379 0.373 0.814

​ Shaded slope 372 0.708 0.464 0.105 0.094 0.817
​ Ratio of SDS/SLS (%) ​ 87.0 80.0 27.7 25.2 100.4

5 Sunlit slope 491 0.811 0.577 0.364 0.352 0.811

​ Shaded slope 868 0.755 0.515 0.118 0.105 0.814
​ Ratio of SDS/SLS (%) ​ 93.1 89.3 32.4 29.8 100.4

6 Sunlit slope 49 0.819 0.585 0.361 0.361 0.818

​ Shaded slope 35 0.810 0.576 0.235 0.225 0.815
​ Ratio of SDS/SLS (%) ​ 98.9 98.5 65.1 62.3 99.6

7 Sunlit slope 169 0.830 0.597 0.397 0.400 0.830

​ Shaded slope 128 0.808 0.574 0.100 0.091 0.815
​ Ratio of SDS/SLS (%) ​ 97.3 96.1 25.2 22.8 98.2

8 Sunlit slope 94 0.831 0.598 0.321 0.309 0.824

​ Shaded slope 55 0.754 0.514 0.149 0.136 0.816
​ Ratio of SDS/SLS (%) ​ 90.7 86.0 46.4 44.0 99.0

9 Sunlit slope 94 0.795 0.559 0.317 0.303 0.795

​ Shaded slope 111 0.756 0.516 0.164 0.149 0.814
​ Ratio of SDS/SLS (%) ​ 95.1 92.3 51.7 49.2 102.4

10 Sunlit slope 482 0.826 0.593 0.389 0.381 0.826

​ Shaded slope 198 0.739 0.498 0.198 0.183 0.770
​ Ratio of SDS/SLS (%) ​ 89.5 84.0 50.9 48.0 93.2

11 Sunlit slope 210 0.804 0.569 0.348 0.337 0.804

​ Shaded slope 300 0.735 0.493 0.162 0.148 0.805
​ Ratio of SDS/SLS (%) ​ 91.4 86.6 46.6 43.9 100.1
​ Average SDS/SLS (%) ​ 92.7 88.8 43.2 40.6 99.1

between shaded and sunlit slopes, achieving closer alignment of these Fig. 5i to l show the overlay of dSEVI with each of the other four dVIs,
values between the slopes. using the red-green differencing method, with the red channel display­
ing dSEVI and the green channel displaying one other dVI. When the
3.2.2. Extraction of burned forest areas extraction extents of both are consistent, they appear yellow because the
Burned forest areas were mapped using pre-fire GF-1 and post-fire color mixing result of red and green is yellow. When the extraction ex­
GF-6 images, which have a resolution of 16 m and, therefore, should tents differ, they appear either red (representing dSEVI) or green
provide greater detail than the 30 m resolution of Landsat imagery. We (denoting the other dVI). Compared with dSEVI, the other four dVIs
employed five dVIs (dSEVI, dNDVI, dkNDVI, dSAVI, and dEVI) to extract primarily exhibit omission of burned forest areas on shady slopes, as
burned forest areas based on the contrast between the VIs before and indicated by many red patches along shady slopes in these figures,
after the fire. particularly prominent in dSAVI and dEVI.
Fig. 5 shows the original pre- and post-fire GF images (Fig. 5a and b) Fig. 5m to x present a comparison of the extraction effectiveness of
along with the burned forest areas extracted using the five dVIs. In the the five dVIs on two shaded slopes. Fig. 5m and s are the CBERS-04A’s
March 29 post-fire GF-6 image (Fig. 5b), a large burned scar is evident, high-resolution 2 m fusion images of the two slopes. The effectiveness of
characterized by a charcoal black color. Fig. 5c depicts the overlay of dSEVI in extracting burned areas on shaded slopes is evident (Fig. 5n and
pre- and post-fire images using red and green channels, respectively, t, where red represents burned areas and black denotes non-burned
highlighting burned areas in red. Observations from Fig. 5d to h reveal areas). Conversely, dNDVI and dkNDVI exhibit omission in deep
differences in the forest burned areas extracted by the five dVIs. The shadows, while dSAVI and dEVI inaccurately classify almost all shaded
influence of shady slopes is apparent, with some dVIs exhibiting sig­ slope areas as black-colored, non-burned areas.
nificant omission errors, particularly dEVI (Fig. 5h), which extracts only Using CBERS-04A’s 2 m high-resolution color fusion image as a
11,459 ha of burned area, notably less than dSEVI (19,399 ha) (Table 3). reference, 300 sample points were randomly selected to validate the

8
H. Xu et al. International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

Fig. 5. Images series showing various burned forest areas. (a) Pre-fire GF-1 image (March 11), (b) Post-fire GF-6 image (March 29), (c) Overlay of pre- and post-fire
images, (d-h) Extracted burned forest areas using dSEVI, dNDVI, dkNDVI, dSAVI, and dEVI, respectively, (i-l) Overlay of burned forest areas extracted by dSEVI with
the other four dVIs, respectively, using the red-green differencing method, (m, s) CBERS-04A high-resolution 2 m fusion images showing fire burn scars on both slope
sides of ridges, respectively, (n to r and t to x) Burn scars extracted by the five dVIs, where red represents burned areas and black denotes non-burned areas, (n and t):
dSEVI, (o and u): dNDVI, (p and v): dkNDVI, (q and w): dSAVI, and (r and x): dEVI. (For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)

accuracy of the extracted burned forest areas (Table 3). It is found that 3.2.3. Extraction of total burned areas
dSEVI achieves the highest accuracy, with an overall accuracy of 94.3 % The dVIs can extract burned forest areas as indicated above but are
and a Kappa of 0.88, followed by dNDVI, dkNDVI, dSAVI, and dEVI. not capable of identifying burned areas that were not forests in both pre-
Differences among the indices primarily manifest in the omission rates and post-fire events due to limited vegetation covers detectable by VIs in
of the burned areas. Notably, dSEVI exhibits the lowest omission rate, non-vegetated areas. Therefore, we employed the Random Forest (RF) to
effectively addressing the issue of shadow-induced VI suppression. classify the post-fire GF-6 image for extracting the total burned areas,
Conversely, the other dVIs display noticeable problems, with dEVI encompassing both forested and non-forested terrain. Fig. 6a illustrates
reaching an omission error of 51.3 %. the overlay of the RF-extracted total burned areas and the dSEVI-
extracted burned forest areas using the red-green differencing method

9
H. Xu et al. International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

Table 3
Accuracy validation of burned area extraction.
Burned forest Total burned Overall accuracy Kappa Omission (%) Commission (%)
area (ha) area (ha) (%)

dSEVI 19,266 ​ 94.3 0.88 2.7 6.2

dNDVI 18,081 ​ 87.5 0.74 16.0 4.8
dkNDVI 16,961 ​ 81.6 0.63 24.6 6.6
dSAVI 13,221 ​ 72.5 0.48 41.2 5.2
dEVI 11,459 ​ 66.3 0.38 51.3 5.5
RF ​ 23,588 88.0 0.74 4.3 13.5

Fig. 6. (a) Overlay of the RF-extracted total burned areas and the dSEVI-extracted burned forest areas using the red-green method for comparison (the yellow areas
represent burned areas extracted by both methods and the red-colored areas depict burned areas extracted solely by RF). (b) Pre-fire GF-1 image, highlighting a
prominent soil exposure surrounded by forests. (c) Post-fire GF-6 image showing a charcoal-black hue in the exposed soil area after the fire. (d) Overlay of burned
areas extracted by RF and dSEVI in the corresponding soil exposure area, (e) Post-fire Landsat-9 image (March 23), and (f) NBR-extracted total burned area from the
Landsat-9 image (white patches represent burned areas). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of
this article.)

for comparison. The yellow areas represent burned areas extracted by changes after the fire. Burn severity often exhibits heterogeneity within
both methods. The red-colored areas depict burned areas extracted the extent of a fire-affected area, particularly in forested landscapes
solely by RF, which were devoid of forest cover both before and after the (Gibson et al., 2020).
fire. Therefore, the burned areas extracted by RF include both burned Differenced spectral indices like dNBR and dNDVI have been used to
forest and non-forest areas, i.e., the total burned area. The total burned estimate the severity of the burn and the extent of burned areas (Chen
area extracted by RF is 23,588 ha, which is greater than the forest et al., 2022). Burn severity can be identified based on these indices, with
burned area (19,266 ha) extracted by dSEVI (Table 3). Fig. 6b is a pre- higher values indicating greater severity (Epting et al., 2005). Therefore,
fire GF-1 sub-image depicting a large soil exposure surrounded by for­ we equally divided the dSEVI values into three burn severity levels (low,
ests. Fig. 6c displays the post-fire GF-6 sub-image showing a charcoal- moderate, and high) (Fig. 7) to generate a burn severity map (Fig. 8a).
black hue in the corresponding bare soil area after extensive burning. Furthermore, we overlaid the dSEVI image with DEM and slope data and
Fig. 6d shows the overlay of the RF-extracted burned area and the found that the elevation of the fire-affected areas ranged from 2521 to
dSEVI-extracted burned area using the red-green differencing method, 4552 m, with an average elevation of 3494 m. The average slope was
highlighting the burned soil area in red, indicating that the exposed soil 31.2◦ , with a maximum of 76◦ . The distribution of the three burn
area was extracted solely by RF but not by dSEVI. Similarly, the dNBR severity levels—high, moderate, and low—was observed at average el­
method can only extract the total burned area and cannot differentiate evations of 3472 m, 3497 m, and 3505 m, suggesting a transition from
forest fire areas within it (Fig. 6e and 6f). high to low severity as elevation rises. Analysis of Variance (ANOVA)
reveals an F-value of 390.63, which is considerably greater than the F-
critical value of 2.99, and a p-value less than 0.01. This indicates a
3.3. Burn severity statistically significant difference between the groups at these three
average elevation levels. However, there was no significant difference in
An essential metric used in remote sensing for fire assessments is slope distribution among the three levels, as depicted in Fig. 8b, showing
directly linked to the severity of burning. In this study, burn severity is relatively consistent slopes in the mountainous terrain where the fire
described as the degree of spectral changes caused by the burn according occurred. Overall, high elevation and steep terrain pose significant
to the definition of Chatzopoulos-Vouzoglanis et al. (2024), Keeley challenges to fire suppression efforts, prolonging the forest fire for 13
(2009), and Rogan and Yool (2001), which emphasizes the spectral

10
H. Xu et al. International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

Fig. 7. Three-level burn severity map. (a, c) Original post-fire images showing heterogeneous burned areas characterized by three different burn severity levels
marked as 1 (low), 2 (moderate), and 3 (high), respectively, and (b, d) Burn severity maps corresponding to (a) and (c), where blue, green, and red colors denote low,
moderate, and high levels of burn severity. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 8. (a) Burn severity map and (b) 3D terrain map showing fire burn scars.

days. (Fig. 9c) to 0.538 on the March 23 post-fire RSEI map (Fig. 9d),
reflecting a 15 % decline in ecological conditions in response to the
3.4. Impact of the fire on regional ecological conditions Yajiang forest fire in the region.
Fig. 10a illustrates the contrasting trends of RSEI and dSEVI values in
Pre- and post-fire ecological changes of the fire site and its sur­ a traversing profile across the fire site, indicating a negative relationship
rounding area were assessed using RSEI based on Landsat 8/9 images between the two. The profile shows that dSEVI is high at the fire site,
acquired on January 11 (Fig. 9a) and March 23 (Fig. 9b), which are the while RSEI is low. When moving away from the fire site, dSEVI declines,
only two Landsat images available in this fire duration. Since there were whereas RSEI gradually increases, suggesting a gradual reduction in the
no major fires on or after March 23 as analyzed in Section 3.1, the March fire’s impact on the surrounding ecosystem.
23 image can be used to represent the post-fire short-term ecological We created buffer zones every 100 m outside the fire site. Fig. 10b
conditions. shows the mean value of dRSEI (expressed as postRSEI – preRSEI) in­
Fig. 9b illustrates that the Yajiang major forest fire had severe creases across 200 buffer zones spanning a distance of 20 km, with the
ecological effects on the region, as evidenced by a large red patch in the most rapid increase occurring within the first and second buffer zones
center of the image due to the substantial removal of aboveground (100 m and 200 m). This indicates that as distance from the fire site
vegetation by the intense fire. Consequently, the post-fire ecological increases, the dRSEI values generally increase and approach 0, implying
condition map derived by the RSEI technique (Fig. 9d) shows a corre­ a gradual decline in the fire’s impact on the ecosystem and a return to
sponding red patch in the central area, indicating significant ecological regional ecological conditions similar to those before the fire.
degradation at the site. The statistical analysis reveals that the mean
RSEI value decreased from 0.633 on the January 11 pre-fire RSEI map

Fig. 9. RSEI images illustrating ecological changes before and after the fire. (a) Pre-fire Landsat-8 image (01-11-2024, RGB = MIR2, NIR, Green bands), (b) Post-fire
Landsat-9 image (03-23-2024, enhanced by a thermal infrared band), (c) Pre-fire RSEI image derived from (a), and (d) Post-fire RSEI image derived from (b). (For
interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

11
H. Xu et al. International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

Fig. 10. (a) A cross profile showing the contrasting trends of RSEI and dSEVI, indicating a negative relationship between the two, and (b) Regression analysis of the
relationship between dRSEI and the distance from the fire site to the buffer zones. The yellow ring in the left image represents the outer boundary of the buffer zones.

4. Discussion successfully mitigates the influence of shadows on VI values on shady

slopes. Consequently, the computation of dSEVI between pre- and post-
4.1. Impact of topographic shadow fire images can minimize differences in VI values between sunlit and
shaded slopes, thus preventing the omission of forest fire scars and
VIs such as NDVI and SAVI, which rely mainly on the NIR band, are improving the accuracy of burned forest area extraction. The principle
generally considered less effective than NBR because they lack the MIR behind SEVI involves introducing a shadow enhancement factor through
band (van Wagtendonk et al., 2004; Chen et al., 2020). Nevertheless, regression analysis to offset index differences between shaded and sunlit
due to the shortage of suitable images with the MIR band, this study had slopes, thereby reducing topographic influence. Among the other four
to map the burned areas primarily using VIs and also found significant used VIs, NDVI can partially mitigate the impact of shadows but is
differences in performance among the VIs employed. Noticeable dis­ ineffective against deep shadows, resulting in lower accuracy in
crepancies in VI values in regions with significant topographical varia­ extracting burned forest areas than SEVI.
tions between sunny and shaded slopes are the primary reasons for these
performance differences (cf. Lyndersen et al., 2016). 4.2. Eco-environmental effects of the fire
Among the five examined VIs, SAVI and EVI were most significantly
affected by shading, with their index values on shaded slopes being less Forest fires often result in significant changes in remotely-sensed
than 45 % of those on sunny slopes. Verbyla et al. (2008) found that sites spectral properties due to the consumption of vegetation, destruction
with steep north-facing aspects faced away from incoming solar radia­ of leaf chlorophyll, exposure of soil, and modifications in moisture levels
tion and hence had consistently lower pre-fire NBR compared to other of ground surfaces (Epting et al., 2005). These changes can be effectively
aspects. Sesnie et al. (2012) compared the impacts of high topographic captured by the greenness, wetness, dryness, and hotness indicators
relief on EVI and NDVI. Their results demonstrated a negative correla­ integrated within RSEI. Through the analysis of pre- and post-fire
tion between the sun elevation angle and EVI or NDVI values derived changes reflected in these indicators, RSEI can provide estimates of
from Landsat TM data in areas with steep terrain. Matsushita et al. the greenness and wetness loss caused by the fire, as well as the extent of
(2007) indicated that the soil adjustment factor (L) in EVI rendered it soil exposure. Consequently, RSEI reveals a 15 % decrease in regional
more sensitive to topographic conditions than NDVI. This is consistent ecological quality compared to pre-fire conditions, indicating a pro­
with the present study, where the EVI and SAVI values, both also nounced impact of the fire on the ecosystem.
incorporating the L factor, exhibited more significant decay on shaded
slopes than the other VIs used. 4.3. Burned area extraction
The SEVI developed in this study effectively enhances VI values on
shaded slopes and thus reduces the gap in the values between shaded This study demonstrates that the dSEVI can effectively estimate
and sunlit slopes. The introduction of the adjustment factor, δ, burned forest areas. However, when the land surface is predominantly

12
H. Xu et al. International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

bare soil, dSEVI is not suitable for detecting such burned areas because Baig, M.H.A., Zhang, L.F., Shuai, T., Tong, Q.X., 2014. Derivation of a tasselled cap
transformation based on Landsat 8 at-satellite reflectance. Remote Sens. Lett. 5,
the lack of evident vegetation biophysical parameters leads to insignif­
423–431.
icant changes in the index before and after the fire. However, the RF and Belgiu, M., Drăguţ, L., 2016. Random forest in remote sensing: a review of applications
NBR approaches can identify all burned areas, including those with bare and future directions. ISPRS J. Photogramm. Remote Sens. 114, 4–31.
soil. Nevertheless, neither RF nor NBR alone can differentiate between Bretreger, D., Hancock, G.R., Lowry, J., Senanayake, I.P., Yeo, I.Y., 2024. The impacts of
burn severity and frequency on erosion in Western Arnhem Land, Australia. Sensors
burned forest areas and burned bare-soil areas. Therefore, for mapping 24, 2282.
burned areas in mountainous regions, we recommend a combination of Camps-Valls, G., Campos-Taberner, M., Moreno-Martínez, Á., Walther, S., Duveiller, G.,
these two methods. This approach allows for the estimation of the total Cescatti, A., Mahecha, M.D., Muñoz-Marí, J., García-Haro, F.J., Guanter, L.,
Jung, M., Gamon, J.A., Reichstein, M., Running, S.W., 2021. A unified vegetation
burned area while also distinguishing burned forest regions within it. index for quantifying the terrestrial biosphere. Sci. Adv. 7.
Chatzopoulos-Vouzoglanis, K., Reinke, K.J., Soto-Berelov, M., Jones, S.D., 2024. Are fire
5. Conclusions intensity and burn severity associated? Advancing our understanding of FRP and
NBR metrics from Himawari-8/9 and Sentinel-2. Int. J. Appl. Earth Obs. Geoinf. 127,
103673.
This study employed multi-platform remote sensing data to monitor Chen, D., Loboda, T.V., Hall, J.V., 2020. A systematic evaluation of influence of image
and evaluate the Yajiang forest fire, which occurred on March 15, 2024, selection process on remote sensing-based burn severity indices in North American
boreal forest and tundra ecosystems. ISPRS J. Photogramm. Remote Sens. 159,
in southwestern China. By integrating high temporal resolution MODIS 63–77.
and Black Marble NTL images, this research successfully tracked the Chu, T., Guo, X., 2014. Remote sensing techniques in monitoring post-fire effects and
entire fire progression and revealed that the peak of the major fire patterns of forest recovery in boreal forest regions: a review. Remote Sens. 6,
470–520.
occurred between March 16 and 18.
Collins, L., McCarthy, G., Mellor, A., Newell, G., Smith, L., 2020. Training data
The novel SEVI effectively enhanced VI values on shaded slopes, requirements for fire severity mapping using Landsat imagery and random forest.
reducing their disparity between shaded and sunlit areas. Consequently, Remote Sens. Environ. 245, 111839.
Crist, E.P., 1985. A TM tasseled cap equivalent transformation for reflectance factor data.
dSEVI outperformed the other four dVIs in extracting burned forest
Remote Sens. Environ. 17, 301–306.
areas. The combination of dSEVI and RF methods accurately delineated de Carvalho, O.A., Guimaraes, R.F., Silva, C.R., Gomes, R.A.T., 2015. Standardized time-
total burned areas and forest burned areas, revealing a total burned area series and interannual phenological deviation: new techniques for burned-area
of 23,588 ha, with the forest fire area covering 19,266 ha caused by the detection using long-term MODIS-NBR dataset. Remote Sens. 7, 6950–6985.
Epting, J., Verbyla, D., Sorbel, B., 2005. Evaluation of remotely sensed indices for
Yajiang forest fire. assessing burn severity in interior Alaska using Landsat TM and ETM+. Remote Sens.
The fire significantly impacted regional ecological status, resulting in Environ. 96, 328–339.
a 15 % decline in ecological conditions as measured by RSEI values. The Firozjaei, M.K., Kiavarz, M., Homaee, M., Arsanjani, J.J., Alavipanah, S.K., 2021. A novel
method to quantify urban surface ecological poorness zone: a case study of several
use of RSEI to measure the fire’s impact on eco-environments demon­ European cities. Sci. Total Environ. 757, 143755.
strates its potential to be a predictor of ecosystem responses to fires. Gibson, R., Danaher, T., Hehir, W., Collins, L., 2020. A remote sensing approach to
These findings underscore the importance of remote sensing tech­ mapping fire severity in south-eastern Australia using sentinel 2 and random forest.
Remote Sens. Environ. 240, 111702.
nologies in proactive fire management and informed ecological resto­ Giddey, B.L., Baard, J.A., Kraaij, T., 2022. Verification of the differenced normalised
ration. The insights gained from this research provide valuable guidance burn ratio (dNBR) as an index of fire severity in Afrotemperate forest. S. Afr. J. Bot.
for future conservation efforts and assist land managers in making better 146, 348–353.
Gislason, P.O., Benediktsson, J.A., Sveinsson, J.R., 2006. Random forests for land cover
decisions for post-fire rehabilitation.
classification. Pattern Recogn. Lett. 27, 294–300.
Hao, X.Y., Liu, J.X., Heiskanen, J., Maeda, E.E., Gao, S., Li, X.C., 2023. A robust gap-
CRediT authorship contribution statement filling method for predicting missing observations in daily Black Marble nighttime
light data. Gisci. Remote Sens. 60.
Henry, M.C., 2008. Comparison of single- and multi-date Landsat data for mapping
Hanqiu Xu: Writing – original draft, Visualization, Supervision, wildfire scars in Ocala National Forest, Florida. Photogramm. Eng. Remote Sens. 74,
Project administration, Methodology, Funding acquisition, Formal 881–891.
Hu, X., Ban, Y., Nascetti, A., 2021. Uni-temporal multispectral imagery for burned area
analysis, Conceptualization. Jiahui Chen: Validation, Formal analysis,
mapping with deep learning. Remote Sens. 13, 1509.
Data curation. Guojin He: Resources, Data curation. Zhongli Lin: Hu, X.S., Xu, H.Q., 2019. A new remote sensing index based on the pressure-state-
Methodology, Data curation. Yafen Bai: Data curation. Mengjie Ren: response framework to assess regional ecological change. Environ. Sci. Pollut. Res.
Data curation. Hao Zhang: Investigation. Huimin Yin: Data curation. 26, 5381–5393.
Huete, A.R., 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25,
Fenfen Liu: Data curation. 295–309.
Huete, A.R., Didan, K., Miura, T., Rodriguez, E.P., Gao, X., Ferreira, L.G., 2002. Overview
of the radiometric and biophysical performance of the MODIS vegetation indices.
Declaration of competing interest
Remote Sens. Environ. 83, 195–213.
Hultquist, C., Chen, G., Zhao, K.G., 2014. A comparison of Gaussian process regression,
The authors declare that they have no known competing financial random forests and support vector regression for burn severity assessment in
interests or personal relationships that could have appeared to influence diseased forests. Remote Sens. Lett. 5, 723–732.
Jensen, J.R., 2000. Remote sensing of environment: an earth resource perspective, 1st ed.
the work reported in this paper. Prentice Hall, Upper Saddle River, N.J.
Kamran, M., Yamamoto, K., 2023. Evolution and use of remote sensing in ecological
Data availability vulnerability assessment: a review. Ecol. Indic. 148, 110099.
Keeley, J.E., 2009. Fire intensity, fire severity and burn severity: a brief review and
suggested usage. Int. J. Wildland Fire 18, 116–126.
Data will be made available on request. Key, C.H., Benson N.C., 2001. The normalized burn ratio (NBR): a Landsat TM
radiometric measure of burn severity. [Link]
(Accessed 19 May 2024).
Acknowledgment Knopp, L., Wieland, M., Rättich, M., Martinis, S., 2020. A deep learning approach for
burned area segmentation with Sentinel-2 data. Remote Sens. 12, 2422.
This work was supported by the National Natural Science Foundation Kurbanov, E., Vorobev, O., Lezhnin, S., Sha, J., Wang, J., Li, X., Cole, J., Dergunov, D.,
Wang, Y., 2022. Remote sensing of forest burnt area, burn severity, and post-fire
of China [31971639]. recovery: a review. Remote Sens. 14, 4714.
Lawrence, R.L., Wood, S.D., Sheley, R.L., 2006. Mapping invasive plants using
References hyperspectral imagery and Breiman Cutler classifications (Random Forest). Remote
Sens. Environ. 100, 356–362.
Lee, R.J., Chow, T.E., 2015. Post-wildfire assessment of vegetation regeneration in
Alcaras, E., Costantino, D., Guastaferro, F., Parente, C., Pepe, M., 2022. Normalized burn
Bastrop, Texas, using Landsat imagery. Gisci. Remote Sens. 52, 609–626.
ratio plus (NBR+): a new index for Sentinel-2 imagery. Remote Sens. 14, 1727.
Lee, K., Kim, B., Park, S., 2023. Evaluating the potential of burn severity mapping and
Badhan, M., Shamsaei, K., Ebrahimian, H., Bebis, G., Lareau, N.P., Rowell, E., 2024. Deep
transferability of Copernicus EMS data using Sentinel-2 imagery and machine
learning approach to improve spatial resolution of GOES-17 wildfire boundaries
learning approaches. Gisci. Remote Sens. 60, 2192157.
using VIIRS satellite data. Remote Sens. 16, 715.

13
H. Xu et al. International Journal of Applied Earth Observation and Geoinformation 134 (2024) 104210

Lentile, L.B., Holden, Z.A., Smith, A.M.S., Falkowski, M.J., Hudak, A.T., Morgan, P., Masuoka, E.J., 2018. NASA’s Black Marble nighttime lights product suite. Remote
Lewis, S.A., Gessler, P.E., Benson, N.C., 2006. Remote sensing techniques to assess Sens. Environ. 210, 113–143.
active fire characteristics and post-fire effects. Int. J. Wildland Fire 15, 319–345. Rouse, J.W., Haas, R.H., Schell, J.A., Deering, D.W., 1973. Monitoring vegetation
Lyndersen, J.M., Collins, B.M., Miller, J.D., Fry, D.L., Stephens, S.L., 2016. Relating fire systems in the Great Plains with ERTS. In: Conference Proceedings of the Third ERTS
caused change in forest structure to remotely sensed estimates of fire severity. Fire Symposium, NASA SP-351, Washington D.C., 309–317.
Ecol. 12, 99–116. Sesnie, S.E., Dickson, B.G., Rosenstock, S.S., Rundall, J.M., 2012. A comparison of
Maity, S., Das, S., Pattanayak, J.M., Bera, B., Shit, P.K., 2022. Assessment of ecological Landsat TM and MODIS vegetation indices for estimating forage phenology in desert
environment quality in Kolkata urban agglomeration, India. Urban Ecosystems 25, bighorn sheep (Ovis canadensis nelsoni) habitat in the Sonoran Desert, USA. Int. J.
1137–1154. Remote Sens. 33, 276–286.
Mallinis, G., Mitsopoulos, I., Chrysafi, I., 2018. Evaluating and comparing Sentinel 2A Shi, T.T., Xu, H.Q., 2019. Derivation of tasseled cap transformation coefficients for
and Landsat-8 Operational Land Imager (OLI) spectral indices for estimating fire Sentinel-2 MSI at-Sensor reflectance data. IEEE J. Sel. Top. Appl. Earth Obs. Remote
severity in a Mediterranean pine ecosystem of Greece. Gisci. Remote Sens. 55, 1–18. Sens. 12, 4038–4048.
Masek, J.G., Huang, C., Wolfe, R., Cohen, W., Hall, F., Kutler, J., Nelson, P., 2008. North Smith, C.W., Panda, S.K., Bhatt, U.S., Meyer, F.J., Badola, A., Hrobak, J.L., 2021.
American forest disturbance mapped from a decadal Landsat record, Remote Sens. Assessing wildfire burn severity and its relationship with environmental factors: a
Environ. 112, 2914–2926. case study in Interior Alaska boreal forest. Remote Sens. 13, 1966.
Matsushita, B., Yang, W., Chen, J., Onda, Y., Qiu, G., 2007. Sensitivity of the enhanced Soverel, N.O., Coops, N.C., Perrakis, D.D.B., Daniels, L.D., Gergel, S.E., 2011. The
vegetation index (EVI) and normalized difference vegetation index (NDVI) to transferability of a dNBR-derived model to predict burn severity across 10 wildland
topographic effects: a case study in high-density cypress forest. Sensors 7, fires in western Canada. Int. J. Wildland Fire 20, 518–531.
2636–2651. Stow, D., Petersen, A., Rogan, J., Franklin, J., 2007. Mapping burn severity of
Mellor, A., Boukir, S., Haywood, A., Jones, S., 2015. Exploring issues of training data mediterranean-type vegetation using satellite multispectral data. Gisci. Remote Sens.
imbalance and mislabelling on random forest performance for large area land cover 44, 1–23.
classification using the ensemble margin. ISPRS J. Photogramm. Remote Sens. 105, van Wagtendonk, J.W., Root, R.R., Key, C.C., 2004. Comparison of AVIRIS and Landsat
155–168. ETM+ detection capabilities for burn severity. Remote Sens. Environ. 92, 397–408.
Miller, C.W., Harvey, B.J., Kane, V.R., Moskal, L.M., Alvarado, E., 2023. Different Verbyla, D.L., Kasischke, E.S., Hoy, E.E., 2008. Seasonal and topographic effects on
approaches make comparing studies of burn severity challenging: a review of estimating fire severity from Landsat TM/ETM+ data. Int. J. Wildland Fire 17,
methods used to link remotely sensed data with the composite burn index. Int. J. 527–534.
Wildland Fire 32, 449–475. Weber, K.T., Seefeldt, S., Moffet, C., Norton, J., 2008. Comparing fire severity models
Morresi, D., Marzano, R., Lingua, E., Motta, R., Garbarino, M., 2022. Mapping burn from post-fire and pre/post-fire differenced imagery. Gisci. Remote Sens. 45,
severity in the western Italian Alps through phenologically coherent reflectance 392–405.
composites derived from Sentinel-2 imagery. Remote Sens. Environ. 269, 112800. Wingfield, M.J., Brockerhoff, E.G., Wingfield, B.D., Slippers, B., 2015. Planted forest
Patterson, M.W., Yool, S.R., 1998. Mapping fire-induced vegetation mortality using health: the need for a global strategy. Science 349, 832–836.
Landsat thematic mapper data: a comparison of linear transformation techniques. Xu, H.Q., 2015. Retrieval of the reflectance and land surface temperature of the newly-
Remote Sens. Environ. 65, 132–142. launched Landsat 8 satellite. Chinese J. Geophys. 58, 741–747.
Qureshi, S., Alavipanah, S.K., Konyushkova, M., Mijani, N., Fathololomi, S., Firozjaei, M. Xu, H.Q., Wang, M.Y., Shi, T.T., Guan, H.D., Fang, C.Y., Lin, Z.L., 2018. Prediction of
K., Homaee, M., Hamzeh, S., Kakroodi, A.A., 2020. A remotely sensed assessment of ecological effects of potential population and impervious surface increases using a
surface ecological change over the Gomishan Wetland. Iran. Remote Sens. 12 (18), remote sensing based ecological index (RSEI). Ecol. Indic. 93, 730–740.
2989. Xu, H.Q., Wang, Y.F., Guan, H.D., Shi, T.T., Hu, X.S., 2019. Detecting ecological changes
Rashkovetsky, D., Mauracher, F., Langer, M., Schmitt, M., 2021. Wildfire detection from with a remote sensing based ecological index (RSEI) produced time series and
multisensor satellite imagery using deep semantic segmentation. IEEE J. Sel. Top. change vector analysis. Remote Sens. 11, 2345.
Appl. Earth Obs. Remote Sens. 14, 7001–7016. Xu, H.Q., Lin, M.J., Wang, Y.F., Guan, H.D., Tang, F., 2024a. Quantitatively exploring the
Rogan, J., Yool, S.R., 2001. Mapping fire-induced vegetation depletion in the Peloncillo influence of geographical conditions on ecological quality using a novel remote
Mountains Arizona and New Mexico. Int. J. Remote Sens. 22, 3101–3121. sensing model: a comparison between two geographical disparity regions in China.
Román, M.O., Wang, Z., Sun, Q., Kalb, V., Miller, S.D., Molthan, A., Schultz, L., Bell, J., Geo-Spat. Inf. Sci. 1–18. [Link]
Stokes, E.C., Pandey, B., Seto, K.C., Hall, D., Oda, T., Wolfe, R.E., Lin, G., Xu, H.Q., Ren, M.J., Lin, M.J., 2024b. Cross-comparison of Landsat-8 and Landsat-9 data:
Golpayegani, N., Devadiga, S., Davidson, C., Sarkar, S., Praderas, C., Schmaltz, J., a three-level approach based on underfly images. Gisci. Remote Sens. 61, 2318071.
Boller, R., Stevens, J., Ramos González, O.M., Padilla, E., Alonso, J., Detrés, Y., Zhai, Y.G., Roy, D.P., Martins, V.S., Zhang, H.K.K., Yan, L., Li, Z.B., 2022. Conterminous
Armstrong, R., Miranda, I., Conte, Y., Marrero, N., MacManus, K., Esch, T., United States Landsat-8 top of atmosphere and surface reflectance tasseled cap
transformation coefficients. Remote Sens. Environ. 274, 112992.

14