A method for constructing a burnt area detection index and a burnt area detection method

Abstract

This invention relates to the field of remote sensing monitoring technology, specifically to a method for constructing a fire-affected area detection index and a method for detecting fire-affected areas. The method includes: a novel red-edge burned area index (REBAI) based on the reflectance of red light, first red edge, second red edge, near-infrared, and short-wave infrared bands from the Sentinel-2 satellite. It is constructed through ratio calculation (quantifying the spectral jump differences in the red-edge band) and normalization calculation (distinguishing the magnitude relationship of reflectance). During construction, the bands are resampled to a 10-meter resolution to ensure spatial consistency. During detection, 16 adaptive threshold segmentation algorithms are used to process the REBAI, and pixels identified by 14 or more algorithms are defined as the final fire-affected areas. This invention significantly improves the detection accuracy and stability of small-area, low-intensity fire-affected areas, and is applicable to various vegetation types such as forests and grasslands, providing reliable technical support for post-fire loss assessment, ecological restoration, and climate change research.

Description

Technical Field

[0001] This invention relates to the field of remote sensing monitoring technology, and more specifically to a method for constructing a fire detection index and a method for detecting fire-affected areas. In particular, it relates to a method for constructing a fire detection index and a method for detecting fire-affected areas based on the Sentinel-2 red-edge band. Background Technology

[0002] Against the backdrop of climate change, the number of burned areas (exposed areas destroyed by fire) is increasing globally. Accurate monitoring of burned areas is crucial for ecological protection and disaster prevention. Fire, as a severe disturbance to ecosystems, not only directly causes vegetation destruction and biodiversity loss, but can also trigger secondary problems such as soil degradation and water erosion, and even exacerbate climate change by releasing greenhouse gases. Therefore, timely and accurate understanding of the scope, severity, and dynamic changes of burned areas is fundamental to developing ecological restoration plans and assessing the long-term ecological impacts of fires. Simultaneously, burned area monitoring data is also a key basis for disaster management, providing important references for secondary disaster early warning and the allocation of rescue resources, minimizing the threat of fire to people's lives and property. Furthermore, in the field of climate change research, burned areas, as important carriers of the interaction between fire and climate, contribute to a deeper understanding of climate feedback mechanisms such as the carbon cycle and surface albedo, providing scientific support for addressing global climate change.

[0003] In the field of fire monitoring, the existing remote sensing monitoring system for fire-affected areas mainly relies on multi-source satellite remote sensing data, spectral index analysis, and machine learning modeling methods.

[0004] Regarding satellite data sources, current satellites and sensors used for monitoring burned areas include the Landsat series satellites, Sentinel series satellites, PlanetScope satellites, and MODIS and VIIRS sensors. Among them:

[0005] With its long-term data series and medium resolution of 30 meters, the Landsat series satellites have become the benchmark data source for dynamic monitoring of burned areas. However, their long revisit period can easily lead to information gaps in the time series of monitoring data, affecting the accurate depiction of the spatiotemporal evolution of burned areas.

[0006] The synthetic aperture radar (SAR) carried by Sentinel-1 is not limited by lighting and weather conditions and can capture changes in surface roughness after vegetation burning through microwave backscattering characteristics. However, its data requires complex processing procedures, which increases the threshold for data application.

[0007] PlanetScope, with its high resolution of 3-5 meters and daily revisit capability, can accurately identify the boundaries of small-scale burned areas, but its data costs are relatively high.

[0008] MODIS sensors have daily revisit capability and are irreplaceable in monitoring fire-prone areas worldwide, but their spatial resolution is low, making them difficult to directly apply to the fine identification of small-scale fires and their edges.

[0009] With its 375-meter resolution, daily global coverage, and excellent fire detection sensitivity, VIIRS has a significant advantage in real-time monitoring of small-scale burned areas. However, its spatial resolution is still lower than Landsat / Sentinel-2, making it difficult to finely characterize the burning edge. Furthermore, it relies on the thermal infrared band and cannot directly assess the changes in spectral characteristics after vegetation burning.

[0010] In comparison, the Sentinel-2 multispectral instrument, with its combination of 13 bands and 10-60 meter resolution, demonstrates unique advantages in the analysis of spectral characteristics of burned areas. Its five-day revisit cycle can monitor the dynamic changes of burned areas, and its red-edge band is highly sensitive to vegetation recovery after fire. It has been widely used for fine identification of scorched areas, assessment of fire severity, and monitoring of recovery. It has formed comprehensive advantages in terms of data openness, multi-dimensional monitoring capabilities, and the breadth of research applications, making it an important data source for current monitoring of burned areas.

[0011] In spectral index analysis, existing studies have extracted burned areas using various spectral indices and conducted comparative analyses. Traditional indices, such as the Normalized Burn Ratio (NBR), quantify the degree of vegetation burning through a combination of near-infrared and short-wave infrared, but their sensitivity to lightly burned areas is insufficient. Vegetation indices, such as the Normalized Difference Vegetation Index (NDVI) and the Soil-Adjusted Vegetation Index (SAVI), combine red and near-infrared bands to extract burned areas from the perspective of vegetation state changes and reduced soil disturbance, but they also struggle to identify low-intensity burned areas. The Burned Area Index (BAI) was designed for burned areas in the Mediterranean ecosystem, improving detection accuracy by enhancing post-fire surface information, but its sensitivity is weak in other areas, especially for small burned areas. The Mid-Infrared Burn Index (MIRBI) accurately defines the boundaries of burned areas by optimizing band combinations, but it suffers from a high false negative rate and difficulty in identifying small burned areas. In summary, these indices generally exhibit low accuracy in detecting lightly burned and small burned areas.

[0012] In machine learning modeling, the Random Forest (RF) algorithm is widely used in burnt area classification due to its anti-overfitting properties and feature importance evaluation capabilities. Support Vector Machines (SVMs) achieve efficient classification of multispectral images through kernel function mapping, offering a cost advantage in rapid mapping scenarios. Deep learning frameworks such as U-Net achieve pixel-level high-precision identification of burnt areas through Convolutional Neural Networks (CNNs), performing particularly well in complex terrain and mixed land cover scenarios. The recently developed Transformer architecture captures spatial contextual information through a self-attention mechanism, achieving improved accuracy compared to traditional CNN models. Although machine learning models significantly improve the accuracy of burnt area extraction through nonlinear modeling capabilities, they still face common challenges such as high data annotation costs, long model training times and complex hyperparameter optimization, and limited model generalization.

[0013] In summary, existing methods for monitoring burned areas still face bottlenecks in terms of accuracy. Specifically, existing satellite data sources each have limitations, making it difficult to simultaneously meet the requirements of high resolution, high revisit frequency, low cost, and ease of processing; existing spectral indices have low accuracy in detecting small-area, low-intensity burned areas and lack adaptability to different regions and vegetation types; while machine learning methods can improve accuracy, they suffer from high data annotation costs, complex training, and poor generalization, resulting in a high application threshold. Therefore, there is an urgent need for a monitoring method with clear physical meaning, capable of accurately identifying small-area, low-intensity burned areas, applicable to different degrees of burn severity, and requiring no complex modeling, to address the aforementioned shortcomings of existing technologies. Summary of the Invention

[0014] In view of this, the present invention provides a method for constructing a fire-affected area detection index and a method for detecting fire-affected areas, aiming to solve the bottleneck problems in the accuracy of existing fire-affected area monitoring methods. Specifically, existing spectral indices have low accuracy in detecting small-area, low-intensity fire-affected areas and insufficient adaptability to different regions and vegetation types. Machine learning methods also suffer from high data labeling costs, complex training, and poor generalization. This invention improves the accuracy and stability of fire-affected area detection, enabling accurate identification of small-area, low-intensity fire-affected areas. At the same time, it ensures that the method has clear physical meaning, good regional adaptability, and broad applicability to vegetation types, providing strong support for post-fire loss assessment and ecosystem monitoring without complex modeling.

[0015] To achieve the above objectives, the present invention adopts the following technical solution:

[0016] A method for constructing a fire-marked area detection index, characterized by comprising the following steps:

[0017] Obtain reflectance information for the red band, first red edge band, second red edge band, near-infrared band, and short-wave infrared band in Sentinel-2 satellite imagery;

[0018] Based on the reflectivity of the above-mentioned bands, a fire-marked area detection index is constructed by combining ratio calculation and normalization calculation.

[0019] The ratio calculation is based on the reflectance difference between the first red edge band and the second red edge band, with the ratio RI = ρ_B6 / ρ_B5. The normalization calculation is based on the reflectance difference between the red band, near-infrared band, and short-wave infrared band, with the ratio NDI = (ρ_B8A - ρ_B4 - ρ_B12) / (ρ_B8A + ρ_B4 + ρ_B12). The burnt area detection index is the product of the ratio calculation result and the normalization calculation result, REBAI = RI * NDI, which can highlight the spectral difference between burned areas and unburned vegetation.

[0020] By selecting reflectance data from specific bands of Sentinel-2, and based on the difference in spectral response between burned areas and unburned vegetation, a burnt area detection index is constructed using a combination of ratio calculation and normalization calculation techniques. The aim is to amplify the spectral differences between the two types of areas and lay the foundation for accurate detection.

[0021] In one specific implementation, the red band corresponds to the B4 band of Sentinel-2, the first red edge band corresponds to the B5 band, the second red edge band corresponds to the B6 band, the near-infrared band corresponds to the B8A band, and the short-wave infrared band corresponds to the B12 band.

[0022] By specifying the exact band numbers used to construct the index in Sentinel-2 satellite imagery, and based on the differences in spectral sensitivity of each band to vegetation and burned areas, the operability of the index construction is achieved. The aim is to ensure that the method can be directly implemented using Sentinel-2 data.

[0023] In one specific implementation scheme, the ratio calculation is to calculate the reflectance ratio of the second red-edge band to the first red-edge band to quantify the spectral jump difference in the red-edge band; the normalization calculation is to construct a normalization index that distinguishes the magnitude relationship between the reflectance of the red, near-infrared and short-wave infrared bands.

[0024] By refining the specific methods of the two operations, based on the spectral jump characteristics of the red-edge band and the magnitude relationship of reflectance between the visible and infrared bands, the difference between burned areas and unburned vegetation can be accurately captured, with the aim of improving the index's distinguishing ability.

[0025] In one specific implementation scheme, the fire-affected area detection index is obtained by multiplying the ratio calculation result with the normalization calculation result, thereby amplifying the numerical difference between the fire-affected area and the unburned vegetation and reversing the sign.

[0026] By multiplying the results of the two calculations, based on the technical concept of numerical amplification and sign differentiation, the numerical difference between burned areas and unburned vegetation is maximized, with the aim of more intuitively identifying target areas.

[0027] In one specific implementation, after obtaining the reflectivity information, the method also includes resampling all bands of Sentinel-2 to unify the spatial resolution to 10 meters.

[0028] By resampling satellite image bands, matching of data from different bands is achieved based on the principle of spatial resolution consistency, with the aim of avoiding the impact of resolution differences on the accuracy of index calculation.

[0029] A method for detecting burned areas, characterized by comprising the following steps:

[0030] Acquire Sentinel-2 images of the area to be detected, and calculate the fire detection index according to any of the methods described above;

[0031] The index was processed using 16 adaptive threshold segmentation algorithms, yielding several preliminary identification results.

[0032] Spatial overlay analysis is performed on the preliminary results, and pixels that are identified as burned areas by at least a preset number of algorithms are designated as the final region.

[0033] By combining the aforementioned index construction method with multi-algorithm threshold segmentation technology, and based on the concept of multi-result collaborative verification, the detection of burned areas is achieved, with the aim of reducing the error of a single algorithm and improving the reliability of the results.

[0034] In one specific implementation, the adaptive threshold segmentation algorithm encompasses:

[0035] Statistical distribution algorithms specifically include Median, Mean, Otsu's method, and Percentile method.

[0036] Histogram shape analysis algorithms specifically include Concavity analysis, Prewitt and Mendelsohn's Minimum method, Triangle method, and Zhenzhou's method;

[0037] Information entropy and fuzzy theory algorithms, specifically including Huang's fuzzy method, Maximum Entropy method, and Minimum Cross Entropy method;

[0038] Iterative optimization and clustering algorithms, specifically including K-Means clustering, Intermeans clustering, IsoData iterative self-organizing data analysis algorithm, and Kittler and Illingworth's Minimum Error;

[0039] Mathematical model-driven algorithms, specifically including the moment-preserving method.

[0040] By selecting a variety of adaptive threshold segmentation algorithms and leveraging the complementary principles of different algorithms, the method can cover different scenarios, thereby enhancing its adaptability.

[0041] In one specific implementation, the preset number is 14, meaning that a pixel is included in the final area only when it is identified as a burned area by 14 or more algorithms.

[0042] By clarifying the judgment criteria for multi-algorithm collaboration and based on the principle of high-proportion consensus, the detection results are rigorously screened, with the aim of reducing the risk of misjudgment and ensuring the rigor of the results.

[0043] In one specific implementation scheme, before calculating the detection index, a step is included to create a vegetation mask based on the bare land index to eliminate bare land areas. The bare land index is calculated based on the reflectance of Sentinel-2 in the blue, red, near-infrared, and short-wave infrared bands. The formula for calculating the bare land index is as follows:

[0044]

[0045] In the formula, , , , These refer to the reflectivity of Sentinel-2 in the B2, B4, B8, and B11 bands, respectively.

[0046] By introducing a bare land index to create a vegetation mask, and based on the principle of spectral difference between bare land and burned land, interference areas are eliminated, with the aim of reducing the impact of bare land on the detection results.

[0047] Compared with existing technologies, the present invention provides a method for constructing a fire-affected area detection index and a fire-affected area detection method for accurate monitoring of fire-affected areas, especially for the identification of small-area and low-intensity fire areas. By utilizing the spectral response differences in the red edge, red light, and shortwave infrared bands of Sentinel-2 satellite imagery, a novel Red Edge Burned Area Index (REBAI) is constructed. This index is based on the reflectance differences between fire-affected areas and normal vegetation in the B4, B5, B6, B8A, and B12 bands. It combines the spectral jump characteristics of the red edge band to design a ratio index (RI) and a three-band normalized index (NDI) and performs combined calculations to maximize the highlighting of the spectral differences between the two types of areas. The REBAI values ​​are processed using 16 adaptive threshold segmentation algorithms, and the final fire-affected area identification result is generated through a multi-algorithm collaborative decision-making strategy, achieving efficient detection of fire-affected areas under different climate zones and vegetation types. This effectively improves the accuracy and stability of fire-affected area detection, with high overall accuracy, precision, recall, and F1 score. Indicators such as score are significantly better than existing common indices; it has the following beneficial effects:

[0048] First, a novel Red Edge Burn Area Index (REBAI) was developed using Sentinel-2 data. Its physical meaning is clear, and by synergistically utilizing the red edge and shortwave infrared bands, it significantly improves the spectral separability between burned areas and unburned vegetation, particularly excelling in detecting small-scale and low-severity fires. This effect stems from the fact that REBAI is constructed based on the differences in spectral responses between burned areas and normal vegetation: normal vegetation exhibits a steep jump in reflectance in the red edge bands (B5, B6), while this characteristic is significantly weakened in burned areas. This difference can be quantified by the ratio of B6 to B5 (RI). Simultaneously, the reflectance of normal vegetation in the B8A band is higher than the sum of the B4 and B12 bands, while the opposite is true for burned areas. Based on this, a three-band Normalized Index (NDI) can effectively distinguish between the two. Combining RI and NDI calculations not only amplifies the numerical difference between normal vegetation and burned areas, but also reverses their signs, thereby maximizing spectral separability. This fine spectral difference capture capability is especially crucial for areas with mild burning and small area.

[0049] Secondly, the detection performance is stable and accurate, with all accuracy indicators closely following the mean distribution, demonstrating excellent applicability in diverse forest and grassland ecosystems worldwide. This is thanks to REBAI's design, which fully considers the adaptability to different geographical and ecological conditions: validation was conducted in 12 forest research areas covering frigid to tropical climate zones across 6 continents, and 4 grassland research areas. Evaluation using JM distance and CIV index showed that the JM distance was consistently greater than 1.9, and the average CIV value was only 0.001, indicating that the separation between burned and unburned vegetation remained excellent in different regions. Simultaneously, the use of 16 adaptive threshold segmentation algorithms and multi-algorithm collaborative decision-making (retaining only pixels identified by at least 14 algorithms) effectively reduced the impact of single-algorithm errors, resulting in overall accuracy (96.60%), precision (92.97%), recall (87.12%), and F1 score (0.89) all stably distributed around the mean, maintaining high detection accuracy in both forest and grassland ecosystems. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0051] Figure 1 This is a schematic diagram showing the geographical locations of forest fire-affected areas. In the diagram, (a)-(l) represent 12 typical forest fire-affected area cases at different latitudes and elevations.

[0052] Figure 2 Spectral curves based on Sentinel-2 images for 12 study areas.

[0053] Figure 3 The values ​​of RI and NDI are based on samples from normal vegetation and burned areas, and the differences between them.

[0054] Figure 3 (a) shows the RI values ​​and ratios for normal vegetation and burned areas.

[0055] Figure 3 (b) shows the NDI values ​​and ratios for normal vegetation and burned areas.

[0056] Figure 3 (c) is the product of the RI and NDI ratios of normal vegetation and burned areas.

[0057] Figure 4 This is a schematic diagram showing the separation between 10 bands and 14 indices.

[0058] Figure 5 The JM distance and CIV values ​​are for the eight indices with high separation.

[0059] Figure 6 The ranking results are based on feature importance.

[0060] Figure 6 (a) The importance ranking of the 24 features in the 12 study areas.

[0061] Figure 6 (b) Mean and standard deviation of the importance ranking of 24 features in 12 study areas.

[0062] Figure 7 The image shows the results of extracting seven indices from burned areas in 12 study regions.

[0063] In the figure, T1 and T2 represent standard false-color images before and during the fire incident, respectively.

[0064] In the figure: 1: Normalized Difference Vegetation Index (NDVI); 2: Enhanced Vegetation Index (EVI); 3: Vegetation Response Index (VRI); 4: Soil-Adjusted Vegetation Index (SAVI); 5: Normalized Burn Ratio (NBR); 6: Normalized Burn Ratio Plus (NBR+); 7: Red-Edge Burned Area Index (REBAI).

[0065] Figure 8 The accuracy of fire-marked areas identification in 7 indices for 12 study areas.

[0066] Figure 8 (a)- Figure 8 (d) The overall accuracy, precision, recall, and F1 score for identifying burned areas are shown in order.

[0067] Figure 9 Comparison of burned areas extracted based on single-phase REBAI (i1) and dual-phase REBAI intensity difference (i2).

[0068] Taking area i in the figure as an example, the area within the red box is the region where the difference in the recognition effect of burned areas is more obvious.

[0069] Figure 10 A comparison of the results of fire-affected areas identified by the threshold method and the random forest method.

[0070] The area within the red box in the image shows the region where the difference in the identification effect of burned areas is more obvious.

[0071] Figure 11 Image showing the results of fire damage identification in grassland cases.

[0072] In the figure, m, n, o, p are true-color composite images of the grassland study area, m1, n1, o1, p1 are REBAI index images, and m2, n2, o2, p2 are images of the extracted burned areas.

[0073] Figure 12 Comparison of the recognition accuracy of REBAI and three types of indices for burned areas. Detailed Implementation

[0074] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] With global climate change leading to increased wildfire frequency, accurate monitoring of burned areas is crucial for assessing forest carbon cycles and guiding ecological restoration efforts. However, traditional methods often struggle to accurately define small or low-intensity burned areas. To address the bottleneck in the accuracy of existing monitoring methods for burned areas, this invention develops a novel Red Edge Burned Area Index (REBAI) using Sentinel-2 data. This index distinguishes burned and unburned forests by the differences in spectral responses in the red edge, red light, and shortwave infrared bands. REBAI is a novel monitoring index composed of multiple spectral bands. The narrow-band characteristics of the red-edge band can accurately capture the degree of damage to the photosynthetic structure of vegetation and the decline in photosynthetic function after a fire. The construction of a new spectral index based on the Sentinel-2 red-edge band has important scientific value. By synergistically utilizing the red-edge and short-wave infrared bands, REBAI improves the spectral separability between burned areas and unburned vegetation, enabling high-precision identification of small-area burns and making it applicable to burned areas of varying degrees of severity. At the same time, it avoids the high computational resource requirements of complex modeling, opening up a new avenue for establishing a universally applicable burn area monitoring system and is of great significance for fire-affected area detection and disaster assessment.

[0076] This invention provides a method for constructing a fire-marked area detection index, comprising:

[0077] Select the study area;

[0078] This invention selected 12 typical forest fire-affected areas distributed across different latitudes and altitudes on six continents worldwide, covering climatic zones such as frigid, temperate, subtropical, and tropical, as well as a vertical zonation from lowland to high mountain. The selection of study areas fully considered geographical heterogeneity and the representativeness of ecosystems, primarily located in global fire hotspots such as Siberia (Russia), British Columbia (Canada), Oregon (USA), southeastern Australia, Amazonas (Brazil), the Congo Basin, Mbong (Cameroon), and Shanxi (China). Figure 1 As shown in the figure, the blue markers represent forest fire-affected areas, and the green markers represent grassland fire-affected areas.

[0079] These study areas primarily consisted of small fires with a burned area of ​​less than 1 square kilometer, exhibiting varying degrees of fire intensity. Specifically, as shown in Table 1, the burned area in all eight study areas (a, b, e, g, h, i, j, k) was less than 1.07 square kilometers. The degree of fire intensity was as follows: Figure 2 As shown in the spectral curves, the spectral curves of the burned areas in the study areas (f) and (j) are closest to those of the vegetation, indicating the least degree of burning. The spectral curves of the fire reflectance in other areas are different, indicating varying degrees of burning.

[0080] Table 1. Detailed information on the forest fire-affected area study area.

[0081]

[0082] Data source;

[0083] The remote sensing data used in this invention comes from the Sentinel-2 satellite of the European Space Agency's (ESA) Copernicus Program. This satellite system consists of two satellites, Sentinel-2A (launched in 2015) and Sentinel-2B (launched in 2017), operating in a sun-synchronous orbit and providing global coverage multispectral remote sensing imagery. Its onboard multispectral imager can acquire data in 13 spectral bands, covering the visible, near-infrared, and short-wave infrared spectral ranges, with spatial resolutions of 10 meters (blue, green, red, and near-infrared), 20 meters (red edge, narrow near-infrared, and short-wave infrared bands, totaling 6 bands), and 60 meters (aerosols, water vapor, and cirrus clouds). Detailed band information is shown in Table 2. The design of the red edge band (705-753 nm) significantly improves the accuracy of vegetation dynamic monitoring and biophysical parameter retrieval. The satellite has a 5-day global revisit cycle (10 days for a single satellite), supports high spatiotemporal resolution surface observation, and is widely applicable to fields such as land cover classification, agricultural management, forest monitoring, water body identification, and environmental change analysis.

[0084] Table 2. Detailed information on Sentinel-2 data bands.

[0085]

[0086] The data used in this invention comes from Sentinel-2 satellite remote sensing data in the Google Earth Engine (GEE) public archive. The Sentinel-2 data provided by the GEE platform is radiometrically corrected multispectral surface reflectance data, which users can access via ee.ImageCollection("COPERNICUS / S2_SR_HARMONIZED"). The study acquired two images from adjacent dates before and after the event in each study area. The pre-disaster image was used to obtain vegetation cover, while the post-disaster image was used to define the extent of burned areas (as shown in Table 1). It should be noted that the time window setting for the pre-disaster image follows these principles: if valid data exists within three months before the disaster, the nearest temporal image is used first; otherwise, data from the same period of the previous year is automatically used as a substitute to ensure the spatiotemporal comparability of the vegetation baseline.

[0087] Furthermore, this invention employs manual visual interpretation, independently selecting 200 sample points each from burned areas and normal forests in each study area, totaling 4800 sample points, for spectral curve plotting, inter-class separability calculation, and feature variable importance analysis. Based on the same interpretation standard, simultaneous comparative analysis of pre- and post-disaster remote sensing images of burned areas is conducted. High-precision vectorization is used to obtain the true boundary information of burned areas, ultimately generating a reference vector dataset with geospatial attributes, providing a benchmark for subsequent accuracy verification.

[0088] Data preprocessing;

[0089] In the data preprocessing stage, all bands were uniformly resampled to a 10-meter resolution to ensure spatial consistency of the data. In forest fire identification, bare land and fire-damaged land are easily confused due to their similar spectral characteristics. Especially in remote sensing images of specific spectral bands, the high similarity between the two can lead to misclassification when classifying normal vegetation and fire-damaged land. To solve this problem, the study uses the Bare Soil Index (BSI) to create a forest mask to remove bare land areas. Existing studies have shown that the BSI distinction threshold between bare land and vegetation in different areas tends to be close to 0. Therefore, this invention classifies areas with BSI>0 as bare land. The specific process is as follows: First, the BSI is calculated using pre-disaster images, and the calculation formula is shown in Equation (1). Then, the calculation results are binarized with a threshold of 0 to finally generate vegetation mask data.

[0090] (1)

[0091] In the formula, , , , These refer to the reflectivity of Sentinel-2 in the B2, B4, B8, and B11 bands, respectively.

[0092] Construction of a novel fire scar index, REBAI;

[0093] The reflectance peaks and absorption troughs of vegetation spectral curves are formed by the combined effects of the chemical components of leaves (such as chlorophyll, water, and pigments) and cellular structure. The mechanisms by which light interacts with vegetation differ across different wavelength ranges, resulting in characteristic spectral features. After vegetation is burned, biomass decreases and water is lost, causing significant changes in spectral reflectance. For example... Figure 2 As shown, there are significant differences in the reflectance curves between burned areas and normal vegetation. Specifically, compared to normal vegetation, the reflectance curves of burned areas show an increasing trend in the red light band (e.g., the B4 band of Sentinel-2) and both long-wave and short-wave infrared bands (e.g., the B12 band). The former reflects changes in chlorophyll, while the latter reflects changes in vegetation moisture. In the 0.705-0.945 micrometer spectral range, which covers part of the red-edge and near-infrared bands and corresponds to the B5-B9 bands of Sentinel-2, vegetation exhibits high reflectance dominated by cellular structure in these bands, and the reflectance of burned areas is significantly lower than that of normal vegetation. Furthermore, normal vegetation exhibits unique spectral response characteristics in the red-edge band, with a faster increase in reflectance, manifested as a steep slope in reflectance.

[0094] Therefore, based on the difference in reflectance curves between burned areas and normal vegetation, a novel Red-Edge Burned Area Index (REBAI) was constructed. This index was constructed based on the following four aspects:

[0095] (1) By Figure 2 It is evident that there are significant differences in reflectance between normal vegetation and burned areas in the B4, B6, B8A, and B12 bands.

[0096] (2) The vegetation reflectance undergoes a breakthrough jump in the 705-740nm red edge band (i.e., a "critical point" transition from the trough of red light absorption to the near-infrared high reflectance peak). This characteristic can be quantified by the ratio of B6 to B5 (RatioIndex, RI, formula (2)), and this ratio is significantly higher in healthy vegetation than in burned areas. Figure 2 It is evident that the reflectance increase in the B5 to B6 bands is relatively gradual for fire-damaged areas of different degrees of burning, and even a decrease in reflectance is observed in the study area (d). Therefore, the ratio is used instead of the difference B6-B5 to prevent negative values. Figure 3-a shows that the RI values ​​of unburned vegetation range from 2.16 to 3.82, while the RI values ​​of burned areas are all less than 1.40. The RI ratio between the two types of areas is between 1.81 and 3.63.

[0097] (3) In most study areas, the reflectance of normal vegetation exhibits the characteristic of "B4 and B12 bands having the lowest values ​​and B8A band having the highest values," and the reflectance of the B8A band is higher than the sum of the reflectances of the B4 and B12 bands; while the reflectance characteristics of burned areas are the opposite, with the B8A band reflectance being significantly lower than the sum of the reflectances of the B4 and B12 bands. Based on this, this invention designs a three-band Normalized Difference Index (NDI), as shown in formula (3). Figure 3 -b shows that the NDI values ​​of unburned vegetation are all positive, while the NDI values ​​of burned areas are all negative.

[0098] (4) By Figure 3 As can be seen from the -c expression, combining RI and NDI not only widens the gap between unburned and burned vegetation by a factor of 2.56-10.90, but also reverses their signs. Specifically, using... Figure 2 Taking the study areas (f) and (j) with the least severe fire as examples, Figure 3 -a The RI value ratio in region (f) is the lowest at 1.80. When combined with NDI, the difference between normal vegetation and burned areas increases to 2.74 times, with the opposite sign. Similarly, in region (j), after the RI and NDI are combined, the difference between normal vegetation and burned areas further increases to 10.90 times, with the opposite sign.

[0099] The REBAI index, constructed based on the above principles, can highlight the difference between normal vegetation and burned areas to the greatest extent. The smaller the value, the more likely the area is to be a burned area. The specific expressions are shown in equations (2)-(4).

[0100] (2)

[0101] (3)

[0102] (4)

[0103] In the formula, , , , , These are the reflectivities of the B4, B5, B6, B8A, and B12 bands of Sentinel-2, respectively.

[0104] Existing spectral indices;

[0105] To verify REBAI's ability to identify burned areas, this invention selected 13 widely used remote sensing indices for burning areas (Table 3), covering three categories: vegetation response, water sensitivity, and combustion characteristics. Vegetation response indices include the Normalized Difference Vegetation Index (NDVI), Difference Vegetation Index (DVI), Ratio Vegetation Index (RVI), Enhanced Vegetation Index (EVI), Vegetation Response Index (VRI), and Soil-Adjusted Vegetation Index (SAVI). Water sensitivity indices include the Normalized Difference Water Index (NDWI) and the Normalized Difference Moisture Index (NDMI). Combustion characteristic indices include the Normalized Burn Ratio (NBR), Burned Area Index (BAI), Mid-Infrared Burn Index (MIRBI), Normalized Burned Ratio-short wave infrared (NBRSWIR), and Normalized Burn Ratio Plus (NBR+).

[0106] Table 3 Commonly used characteristic indices for monitoring burned areas and their calculation formulas

[0107]

[0108] Rules for identifying pixels in burned areas;

[0109] To verify the effectiveness of the index, this invention employs 16 unsupervised thresholding segmentation algorithms to extract burned areas from target regions. Based on differences in algorithm principles, these algorithms are categorized into five types: statistical distribution, histogram shape analysis, information entropy and fuzzy theory, iterative optimization and clustering, and mathematical model-driven. The statistical distribution category determines the threshold using statistical features such as the image's mean, variance, and median, including Median, Mean, Otsu's, and Percentile. The histogram shape analysis category locates the threshold by analyzing geometric features of the histogram, such as peaks, valleys, concavity, and slope, including Concavity, Prewitt and Mendelsohn's Minimum, Triangle, and Zhenzhou's. The information entropy and fuzzy theory category, based on information theory, achieves segmentation by optimizing information entropy or fuzzy membership, including Huang's fuzzy, Maximum Entropy, and Minimum Cross Entropy. Iterative optimization and clustering methods approximate the optimal solution by iteratively updating class centers or error functions. These methods are suitable for complex distribution scenarios and include K-Means, Intermeans, IsoData, and Kittler and Illingworth's Minimum Error. Mathematical model-driven thresholding methods derive the optimal threshold by constraining the mathematical consistency of the statistical moment features (such as the first three moments) of the image's grayscale histogram before and after binarization. A typical example is the moment-preserving thresholding method proposed by Tsai in 1985.

[0110] For each index image of the study area, bare land areas were first removed using vegetation masks. Then, the corresponding binarized images of burned areas were generated using the 16 threshold segmentation algorithms mentioned above. To further improve the reliability of the results, spatial overlay analysis was performed on the 16 segmentation results for the same study area and the same index. The number of times each pixel was marked as a burned area was counted. Only when a pixel was simultaneously identified as a burned area by at least 14 algorithms was it included in the final result. This multi-algorithm collaborative decision-making strategy effectively reduced the impact of errors from a single threshold segmentation algorithm on the results and significantly enhanced the accuracy and robustness of burned area boundary extraction.

[0111] Performance evaluation; including:

[0112] Spectral separability evaluation: By calculating the Jeffries-Matusita (JM) distance and common isolation value (CIV), the degree of separation between burned areas and unburned vegetation on the REBAI value is quantified, verifying the index's ability to distinguish the spectral differences between the two types of areas and ensuring that it can effectively amplify the spectral differences between them.

[0113] To quantify the ability of indices to distinguish between burned areas and vegetation, the Jeffries-Matusita Distance (JM distance) and the proportion of Common Isolation Values ​​(CIV) were used as separability evaluation indicators. The JM distance measures class separability through the difference in probability distributions, and its calculation formulas are shown in equations (5) and (6). A larger JM value indicates a greater degree of separation between the two classes; a JM greater than 1.9 indicates a better degree of separation. The CIV index quantifies the separability of two classes in the feature space, and its calculation formula is shown in equation (7). A smaller CIV value indicates less overlap in the distributions of the two classes and higher separability; a larger CIV value indicates a higher risk of confusion between classes.

[0114] (5)

[0115] (6)

[0116] (7)

[0117] In the formula, A , B Representing the two categories to be compared, Represents the distance between J and M; and Representing categories and The mean; and Representing categories and Standard deviation; N This represents the total number of intervals that have been divided. i This refers to a range index. , Each refers to an interval i Medium category A and B The frequency.

[0118] Detection accuracy evaluation: Using indicators such as overall accuracy, precision, recall, and F1 score, REBAI's detection results are compared with existing common indices (such as NBR, BAI, etc.) to evaluate REBAI's accuracy in identifying burned areas, with particular attention to its detection performance in small-area, low-intensity burned areas.

[0119] To further verify the importance of the newly developed index REBAI relative to other features, the importance of 10 spectral bands and 14 indices in fire scar identification was compared using a method combining Recursive Feature Elimination with Cross-Validation (RFECV). The core principle of RFE is to achieve feature selection by repeatedly training the model and progressively eliminating the least important features. In each iteration, the algorithm trains a classifier using the current feature set, ranks the features according to a feature importance evaluation metric, and then removes the lowest-ranked features; features removed later have higher importance. Simultaneously, the algorithm uses cross-validation to robustly evaluate the performance of the current feature combination, effectively avoiding overfitting and ensuring that the selected feature subset has good generalization ability.

[0120] In this context, 5-fold cross-validation means randomly dividing the original training dataset into five equal-sized subsets (or "folds"). By rotating the validation set (each fold is used only once as the test set), the performance of the feature subsets is robustly evaluated, effectively mitigating the risk of overfitting while ensuring the selected features possess strong generalization ability. In implementation, this invention uses Random Forest (RF) as the base estimator, employs sequential feature elimination with a step size of 1, and quantifies performance by aggregating classification accuracy scores.

[0121] The Randomization (RF) algorithm constructs decision trees through double randomization: it uses bootstrap sampling with replacement to generate training data for each tree, while simultaneously splitting nodes using random feature subsets. This mechanism significantly reduces inter-tree correlation, improves generalization ability, and prevents overfitting; the final prediction is determined by majority voting. Because of its robustness in handling high-dimensional nonlinear data without regularization, RF has been widely applied in remote sensing image classification and disaster monitoring.

[0122] Adaptability and stability evaluation: REBAI was validated in study areas with different climate zones (arctic to tropical) and different vegetation types (forest, grassland, etc.) around the world. The performance stability of REBAI in diverse scenarios was analyzed, and its applicability across regions and vegetation types was evaluated.

[0123] Using the distribution map of burned areas obtained through visual interpretation and manual digitization as a reference, the results of burned area identification were evaluated using four metrics: overall precision (OA), precision, recall, and F1 score (F1). OA represents the percentage of correctly classified samples out of the total number of samples. Precision represents the proportion of pixels predicted as burned areas that are actually burned areas. Recall is the proportion of pixels that are actually burned areas that are predicted as burned areas. F1 is the harmonic mean of precision and recall, used to comprehensively evaluate precision and recall. The calculation formulas are shown in (8)-(11).

[0124] (8)

[0125] (9)

[0126] (10)

[0127] (11)

[0128] Wherein, TP and TN represent the number of correctly identified burned area pixels and non-burned area pixels, respectively; FP represents the number of pixels predicted as burned area but actually being vegetation; and FN represents the number of pixels predicted as vegetation but actually being burned area.

[0129] Results of comparative analysis of the separability of indices;

[0130] Figure 4 This is the result of a class separability analysis of 10 multispectral bands and 14 remote sensing indices from the Sentinel-2 satellite in 12 study areas. The assessment based on JM distance indicates ( Figure 4 Eight indices—NDVI, EVI, VRI, SAVI, NDMI, NBR, NBR+, and REBAI—exhibited excellent separation with JM distances greater than 1.9 in more than 11 study areas, indicating strong adaptability to different regions and their ability to effectively adapt to environmental differences. In contrast, the remaining indices and spectral bands showed larger fluctuations in JM distance values ​​across different study areas, indicating poor stability in separating burned areas. Further analysis using CIV (Common Isolation Value)... Figure 4 -b) Analysis revealed that the CIV values ​​of the 12 indices NDVI, DVI, RVI, EVI, VRI, SAVI, NDWI, NDMI, NBR, NBRSWIR, NBR+, and REBAI all approached the theoretical optimum of 0 in all 12 study areas, demonstrating good class isolation performance.

[0131] Considering the analysis results of both indicators, the intersection indices that simultaneously satisfy a JM distance greater than 1.9 and a CIV closest to 0 include: NDVI, EVI, VRI, SAVI, NDMI, NBR, NBR+, and REBAI, a total of 8 indices. The specific values ​​of the separation indices for these 8 indices are as follows: Figure 5 As shown.

[0132] Comparison of the importance of the indices;

[0133] This invention employs a recursive feature elimination combined with cross-validation (RFECV) algorithm to rank the importance of 24 features in the extraction of burned areas across 12 study regions with different ecological characteristics. The algorithm iteratively optimizes the feature subset through five-fold cross-validation, ultimately generating independent importance ranking results for each region. Figure 6 The visualization results show that the REBAI index ranks between 1 and 4 in importance across all 12 study regions, with an average ranking of 2.08, significantly outperforming other indices. Notably, the standard deviation of this index across all study regions is only 1.26, indicating that its performance stability is less affected by regional heterogeneity. Further analysis reveals that the indices with an average importance ranking within the top 10 are: REBAI, MIRBI, SAVI, NDVI, NBR+, EVI, NBRSWIR, NDWI, BAI, VRI, DVI, RVI, and NBR, with ranking standard deviations ranging from 1.26 to 6.29.

[0134] Based on the analysis of the separation results in the previous section, six indices, namely NDVI, EVI, VRI, SAVI, NBR, and NBR+, were finally selected to conduct quantitative verification of the accuracy of subsequent fire-marked area identification with REBAI.

[0135] Visualization results of fire-affected area monitoring;

[0136] Based on the fire-marked area identification results of 12 study areas ( Figure 7 A comparative analysis of the detection performance of seven typical vegetation indices revealed significant spatial heterogeneity in their sensitivity to burned areas and their resistance to background interference. As shown in the figure, REBAI demonstrated the best burned area identification performance across all study areas, exhibiting the highest degree of agreement with the burned area extent in true-color images and providing clearer distinction of burned area edges. Figure 7 -(f1-f7) and Figure 7 As can be seen from (k1-k7), REBAI significantly outperforms the other six indices. Figure 7-(j1-j7) shows that REBAI's recognition performance is comparable to the other six indices. NDVI almost failed to identify any burned areas in areas b, f, and l, and identified fewer burned areas in areas c, d, and k, resulting in many false negatives. EVI almost failed to identify any burned areas in area f, and misidentified many burned areas in area k. VRI had some false negatives in area b, missed most burned areas in areas d and f, and had a relatively high number of false negatives in area g. SAVI had some false negatives in area f and missed many burned areas in area k. NBR had false negatives in areas b, e, f, and k. NBR+ had many false negatives in areas c, d, e, f, g, i, and k, and some false negatives in areas b and l.

[0137] Quantitative results in monitoring burned areas;

[0138] The accuracy of fire-marked area identification using seven vegetation indices (NDVI, EVI, VRI, SAVI, NBR, NBR+, and REBAI) was compared across 12 typical study areas (al). The results were quantitatively evaluated using four metrics: overall accuracy (OA), precision, recall, and F1 score. The comparison results are presented in the form of heatmaps. Figure 8 middle.

[0139] In terms of overall accuracy (OA), REBAI showed high accuracy in identifying burned areas across all study areas, maintaining an OA of over 85% in all 12 study areas, with an average OA of 96.60%. Areas a, g, and j had overall accuracy exceeding 99%, and area f was significantly higher than other indices. NDVI showed significant regional differences, with poor overall accuracy in areas b, f, and l, fluctuating between 64.79% and 76.91%. VRI failed in some areas; for example, its overall accuracy in areas f and g was only 74.59% and 68.56%, respectively, indicating limitations in regional adaptability. Looking at individual areas, area f had the lowest overall accuracy (64.30% (NBR+) - 87.91% (REBAI)), while areas a and j had higher overall accuracy (mean above 99%).

[0140] In terms of precision, NBR achieved the best accuracy, followed by REBAI. NBR achieved an average precision of 99.36% across the 12 study regions. Except for region k (56.65%), REBAI's precision ranged from 91.92% to 100% in the other regions. VRI showed significant errors in regions g and k (14.26% / 54.73%), and EVI in region k (37.76%).

[0141] In terms of recall, REBAI had the highest average recall and best stability across all study regions, while most indices showed poor recall stability. REBAI's recall ranged from 71.92% (region f) to 96.50% (region d), with an average recall of 87.12%, while other indices had average recalls below 80%, with NBR+ at only 38.08%. By region, region f had the most severe false negatives, with NBR+ recall as low as 0.39%, followed by region k (10.25%). However, REBAI's recall in region f was significantly higher than other indices, while its recall in region j was comparable to other indices.

[0142] From the F1 score perspective, REBAI showed the best overall performance, with an F1 score ranging from 0.7099 to 0.9527 and an average of 0.8917. It demonstrated the best precision in region a (0.9527) and significantly higher precision in region f than other indices. Other indices showed significant differences in performance across regions. NDVI fluctuated considerably in regions b, f, k, and l (0.0033-0.3195); NBR+ was almost ineffective in all six study regions (0.0078-0.3374); and VRI showed poor precision in regions d, f, and g (0.0719-0.2480). Although NBR exhibited excellent precision, its low recall limited its overall effectiveness. REBAI, by balancing precision and recall, demonstrated stronger robustness in the extraction of burned areas of varying sizes and degrees, making it a preferred index for dynamic monitoring of burned areas.

[0143] This invention validates REBAI against 10 spectral bands and 13 common indices through two separability metrics and feature importance analysis. Subsequently, its accuracy in detecting burned areas was evaluated using 16 adaptive threshold algorithms. In comprehensive validation across 12 diverse global forest study areas, REBAI demonstrated superior performance, particularly excelling in detecting small-scale and low-severity fires, achieving a Jeffries-Matusita (JM) distance >1.9 and a mean common isolation value (CIV) ratio of 0.001, ranking first in feature importance. Detection accuracy is high and stable, with all metrics closely aligned with the mean: overall accuracy (96.60%), precision (92.97%), recall (87.12%), and F1 score (0.89), significantly outperforming common indices (such as NDVI and NBR). Furthermore, REBAI exhibits excellent applicability across different vegetation types, achieving an average F1 score of 0.91 in grassland burned area detection. Its accuracy can be further improved through bi-temporal REBAI analysis or by combining it with machine learning techniques. In summary, REBAI is a physically robust fire-marking index with clear characteristics, providing strong support for post-fire loss assessment and ecosystem monitoring.

[0144] The REBAI index described in this invention is a novel fire-affected area monitoring index with clear physical meaning, applicable to different degrees of fire, and strong regional adaptability. When verifying the applicability of this index, only a simple threshold segmentation method was used, but this index can also have a variety of application scenarios.

[0145] First, this invention only tested the accuracy of extracting burned areas based on REBAI calculated from a single post-disaster image. However, calculating the change intensity of REBAI based on dual-temporal data before and after the fire is expected to further improve the accuracy of burned area identification. Taking test area i as an example (where the original single-phase image F1 score was below 0.9), this invention calculated REBAI based on two remote sensing images of the area before and after the fire, generated a REBAI change intensity map through interpolation, and then combined 16 adaptive threshold segmentation algorithms for multi-strategy burned area extraction. Figure 9 Experimental results show that the overall classification accuracy was significantly improved to 99.45%, the precision reached 98.82%, and the recall rate reached 93.76%, which is 26.70% higher than the single-phase method; the F1 score was improved to 0.9622, an improvement of 13.12%.

[0146] Secondly, this invention only tested the accuracy of fire-affected area extraction based on REBAI combined with a threshold method. REBAI can also be used in conjunction with other indices and machine learning for fire-affected area extraction. Figure 7 and Figure 8 The results show that the accuracy of fire-marked area identification is relatively poor in regions f and k. Therefore, this invention takes test regions f and k as examples, using the top three indices (REBAI, SAVI, NDVI) from the feature separation and importance analysis in sections 3.1 and 3.2 as input features. A random forest classifier is trained using visually interpreted samples as training samples, with 500 trees and other parameters kept at default. Fire-marked area extraction is then performed. The experimental results are as follows: Figure 10 As shown, compared to the REBAI thresholding algorithm ( Figure 10 -f1, k1), using the results of fire-affected areas obtained from random forest classification ( Figure 10 The false negative rate in region f and the false positive rate in region k (f2, k2) are significantly reduced. Accuracy evaluation results show that, compared with the REBAI index and threshold segmentation combination method, the new method improves the F1 score by 5.63% in region f and by 14.72% in region k; the overall accuracy reaches 88.70% in region f and 95.25% in region k.

[0147] Finally, this invention only tested REBAI's effectiveness in identifying forest fire-affected areas, but REBAI can also be used to identify fire-affected areas in other vegetation regions. To verify the transferability of the REBAI index, this invention also selected four grassland fire-affected area study areas located in Oceania, Africa, and South America to test the performance of the novel fire-affected area index in grassland fire-affected area extraction. Detailed information is shown in Table 4. These four study areas are located in Queensland, northeastern Australia, Katanga Province in southern Congo, and Tocantins State in central Brazil, respectively, where fire frequency is relatively high.

[0148] Table 4. Detailed information on the grassland fire-affected areas study region

[0149]

[0150] Test results are as follows Figure 11 As shown in the figure, REBAI can effectively distinguish between grassland and burned areas, and successfully extracts most of the burned areas. In regions m, o, and p, REBAI's burned area extraction results are quite consistent with the actual situation. However, in region n, REBAI has a high false negative rate and a high false positive rate, mainly manifested in misidentifying some riverbanks as burned areas, while omitting some lightly burned areas. The accuracy verification results are shown in Table 5. The REBAI index has high accuracy in identifying burned areas in grassland. The overall accuracy of all four regions exceeds 90%, with average precision, recall, and F1 score of 92.80%, 90.42%, and 0.9138, respectively. Among them, the F1 score of regions m, o, and p is higher than 0.95, while the F1 score of region n is relatively low due to the higher false negative rate. Overall, REBAI also shows good extraction performance for grassland burned areas. Exploring the applicability of novel burnt area indices in grassland ecosystems will not only help assess their cross-regional migration capabilities, but also provide a reference for burnt area monitoring technologies in global grassland areas, and promote the improvement of multi-scenario remote sensing monitoring systems.

[0151] Table 5. Identification accuracy of grassland fire-affected areas

[0152]

[0153] Figure 12By overlaying box plots and violin plots, the accuracy of the REBAI index in identifying burned areas was compared with that of three traditional vegetation indices (vegetation response indices: NDVI / DVI / RVI / EVI / VRI / SAVI; moisture-sensitive indices: NDWI / NDMI; combustion characteristic indices: NBR / BAI / MIRBI / NBRSWIR / NBR+). The figures clearly show that the REBAI index has the narrowest and most concentrated accuracy range, with its mean values ​​for precision, recall, F1 score, and overall accuracy (OA) consistently above 0.8, significantly outperforming the other three indices. Detailed comparative analysis reveals that the REBAI index demonstrates significant advantages in overall performance and stability. First, the F1 score mean in the figure shows that the REBAI index's mean F1 score (0.8917) is 17.42%-48.86% higher than other indices, indicating that it has the best performance in terms of the balance between precision and recall, especially suitable for class imbalance recognition scenarios such as burned areas. Second, all four precision indicators of REBAI fluctuate closely around the mean and median. The distribution range of the violin plot shows that REBAI's overall precision and the 95% confidence interval width of the F1 score are 0.11 and 0.21, respectively, significantly narrower than the other three indices (overall precision and the 95% confidence interval width of the F1 score range are 0.31-0.35 and 0.92-0.95, respectively), indicating that REBAI's precision results have the smallest fluctuation range and the best stability in different study areas.

[0154] The core design principle of vegetation response indices (such as NDVI, EVI, and SAVI) is to utilize the spectral characteristics of vegetation in different wavelength bands. Healthy vegetation exhibits high reflectivity in the near-infrared band (0.77-0.90µm) and strong absorption in the red band (0.62-0.68µm). By constructing a ratio calculation between the near-infrared and red bands, the vegetation signal intensity can be effectively amplified, enabling accurate monitoring of vegetation cover. Moisture sensitivity indices (such as NDMI) use a combination of near-infrared (B8, 0.79-0.90µm) and short-wave infrared band 1 (B11, 1.56-1.66µm). Based on the strong absorption characteristics of water in the short-wave infrared band, a quantitative model of vegetation water content is established, providing key parameters for drought monitoring and crop water stress assessment. Combustion characteristic indices (such as NBR and NBR+) are designed focusing on the spectral differences before and after a fire. Specifically, healthy vegetation maintains high reflectivity in the near-infrared band (B8), while burned surfaces exhibit abnormally high reflectivity in the short-wave infrared band 2 (B12, 2.10-2.28µm) due to vegetation destruction. By quantifying the reflectivity differences between the near-infrared and far-short-wave infrared bands, accurate identification of burned areas can be achieved. However, if... Figure 2As shown, the blue light (B2, 0.45-0.52µm) and green light (B3, 0.54-0.58µm) bands have limited ability to distinguish between vegetation and burned areas. In particular, the reflectance curves of the two bands almost overlap in the green light band, which makes it difficult for the NBR+ index to effectively distinguish between the two types of land cover in most threshold segmentation algorithms. The accuracy of burning area extraction is significantly lower than that of other indices.

[0155] From the perspective of REBAI construction principles, the proposed REBAI first enhances vegetation signals by combining near-infrared (B8A) and red (B4) bands, quantifies vegetation moisture content using far-shortwave infrared (B12), and then introduces red-edge bands (B5, B6, 0.70-0.75µm) to construct ratio calculations, further amplifying the spectral differences between vegetation and burned areas. This band combination strategy fully integrates vegetation physiological characteristics (reflectivity differences) and post-burn surface characteristics (moisture loss). Experimental verification shows that the REBAI index exhibits excellent ability to distinguish burned areas in various threshold segmentation algorithms, significantly improving the accuracy and reliability of burned area monitoring.

[0156] Compared to existing metrics, REBAI performs better in detecting small-scale and lightly burned areas. Specific conclusions are as follows:

[0157] First, REBAI demonstrated excellent inter-class separation ability and high feature importance in all 12 small-scale burned areas with different burn degrees and climate types. Specifically, the JM distance index consistently exceeded 1.9, the average CIV was 0.001, and it ranked first in feature importance on average, significantly outperforming the other 13 commonly used indices.

[0158] Secondly, REBAI performed best in identifying burned areas across all study areas, closely matching the actual situation. When comprehensively evaluating the overall precision, accuracy, recall, and F1 score across the 12 study areas, REBAI showed the best performance. Specifically, REBAI's average overall precision, accuracy, recall, and F1 score across the 12 study areas were 96.60%, 92.97%, 87.12%, and 0.8917, respectively. Compared to the SAVI index, which had the second-best F1 score, REBAI improved the F1 score by 9.11%.

[0159] Furthermore, this invention evaluated the stability of REBAI relative to three vegetation indices. The results showed that REBAI had the highest accuracy and was the most stable, with all accuracy indices exceeding 0.8. In addition, test results demonstrated that using dual-temporal REBAI before and after the fire, as well as incorporating machine learning algorithms, effectively improved the accuracy of fire extraction, proving its superiority as a feature engineering approach.

[0160] Finally, this invention analyzed the performance of REBAI in grassland fire site extraction. The average F1 score of the four grassland fire site cases was above 0.9, indicating that the index has applicability across vegetation types.

[0161] The various embodiments described in this specification are presented in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a fire-marked area detection index, characterized in that, Includes the following steps: Acquire reflectance data for the red band, first red edge band, second red edge band, near-infrared band, and short-wave infrared band of Sentinel-2 satellite imagery; Based on the reflectance of the red band, the first red edge band, the second red edge band, the near-infrared band, and the short-wave infrared band, the fire detection index is obtained through a combination of ratio calculation and normalization calculation; the red band is the B4 band of Sentinel-2 satellite imagery, the first red edge band is the B5 band, the second red edge band is the B6 band, the near-infrared band is the B8A band, and the short-wave infrared band is the B12 band; The ratio calculation is based on the reflectance difference between the first red-edge band and the second red-edge band. The ratio is calculated as RI = ρ_B6 / ρ_B5, where ρ_B6 represents the reflectance of the B6 band and ρ_B5 represents the reflectance of the B5 band. The normalization calculation is based on the reflectance differences between the red band, near-infrared band, and short-wave infrared band. The normalized difference index (NDI) for the outer band is calculated as follows: NDI = (ρ_B8A - ρ_B4 - ρ_B12) / (ρ_B8A + ρ_B4 + ρ_B12), where ρ_B8A represents the reflectance of the B8A band, ρ_B4 represents the reflectance of the B4 band, and ρ_B12 represents the reflectance of the B12 band. The burnt area detection index is the product of the ratio calculation result and the normalization calculation result: REBAI = RI * NDI, which highlights the spectral differences between burned areas and unburned vegetation.

2. The method for constructing a burnt area detection index according to claim 1, characterized in that, The ratio calculation specifically involves: calculating the ratio of the reflectance of the second red edge band to the reflectance of the first red edge band to quantify the spectral jump difference between burned areas and unburned vegetation in the red edge band; the normalization calculation specifically involves: constructing a normalization index that can distinguish the magnitude relationship between the reflectances of the red light band, near-infrared band, and short-wave infrared band.

3. The method for constructing a burnt area detection index according to claim 2, characterized in that, The fire-affected area detection index is obtained by multiplying the ratio calculation result with the normalization calculation result, which amplifies the numerical difference between the fire-affected area and the unburned vegetation and reverses the sign.

4. The method for constructing a fire-marked area detection index according to claim 1, characterized in that, After acquiring the reflectivity data, the process also includes resampling all bands of the Sentinel-2 satellite imagery to unify the spatial resolution of each band to 10 meters.

5. A method for detecting burnt areas, characterized in that, Includes the following steps: Acquire Sentinel-2 satellite imagery of the area to be detected, and calculate the fire detection index of the area to be detected based on the method of any one of claims 1-4; Sixteen adaptive threshold segmentation algorithms were used to process the fire-scar detection index, resulting in several preliminary fire-scar identification results. Spatial overlay analysis is performed on the multiple preliminary fire site identification results, and the pixels that are identified as fire sites by at least a preset number of algorithms are determined as the final fire site areas.

6. The method for detecting burned areas according to claim 5, characterized in that, The adaptive threshold segmentation algorithm includes: Statistical distribution algorithms specifically include Median, Mean, Otsu's method, and Percentile method. Histogram shape analysis algorithms include Concavity analysis, Prewitt and Mendelsohn's Minimum method, Triangle method, and Zhenzhou's method. Information entropy and fuzzy theory algorithms, specifically including Huang's fuzzy method, Maximum Entropy method, and Minimum Cross Entropy method; Iterative optimization and clustering algorithms, specifically including K-Means clustering, Intermeans clustering, IsoData iterative self-organizing data analysis algorithm, and Kittler and Illingworth's Minimum Error; Mathematical model-driven algorithms, specifically including the moment-preserving method.

7. The method for detecting burned areas according to claim 5, characterized in that, The preset number is 14, meaning that a pixel is included in the final burned area only when it is identified as a burned area by 14 or more of the aforementioned adaptive threshold segmentation algorithms.

8. The method for detecting burned areas according to claim 5, characterized in that, Before calculating the burned area detection index, the process includes creating a vegetation mask based on the bare land index to remove bare land areas from the area to be detected. The bare land index is calculated based on the reflectance of the blue, red, near-infrared, and short-wave infrared bands of Sentinel-2 satellite imagery. The formula for calculating the bare land index is as follows: ； In the formula, , , , These refer to the reflectivity of Sentinel-2 in the B2, B4, B8, and B11 bands, respectively.

Images