Development of a Novel Burned-Area Subpixel Mapping (BASM) Workﬂow for Fire Scar Detection at Subpixel Level

: The accurate detection of burned forest area is essential for post-ﬁre management and assessment, and for quantifying carbon budgets. Therefore, it is imperative to map burned areas accurately. Currently, there are few burned-area products around the world. Researchers have mapped burned areas directly at the pixel level that is usually a mixture of burned area and other land cover types. In order to improve the burned area mapping at subpixel level, we proposed a Burned Area Subpixel Mapping (BASM) workﬂow to map burned areas at the subpixel level. We then applied the workﬂow to Sentinel 2 data sets to obtain burned area mapping at subpixel level. In this study, the information of true ﬁre scar was provided by the Department of Emergency Management of Hunan Province, China. To validate the accuracy of the BASM workﬂow for detecting burned areas at the subpixel level, we applied the workﬂow to the Sentinel 2 image data and then compared the detected burned area at subpixel level with in situ measurements at ﬁfteen ﬁre-scar reference sites located in Hunan Province, China. Results show the proposed method generated successfully burned area at the subpixel level. The methods, especially the BASM-Feature Extraction Rule Based (BASM-FERB) method, could minimize misclassiﬁcation and effects due to noise more effectively compared with the BASM-Random Forest (BASM-RF), BASM-Backpropagation Neural Net (BASM-BPNN), BASM-Support Vector Machine (BASM-SVM), and BASM-notra methods. We conducted a comparison study among BASM-FERB, BASM-RF, BASM-BPNN, BASM-SVM, and BASM-notra using ﬁve accuracy evaluation indices, i.e., overall accuracy (OA), user’s accuracy (UA), producer’s accuracy (PA), intersection over union (IoU), and Kappa coefﬁcient (Kappa). The detection accuracy of burned area at the subpixel level by BASM-FERB’s OA, UA, IoU, and Kappa is 98.11%, 81.72%, 74.32%, and 83.98%, respectively, better than BASM-RF’s, BASM-BPNN’s, BASM-SVM’s, and BASM-notra’s, even though BASM-RF’s and BASM-notra’s average PA is higher than BASM-FERB’s, with 89.97%, 91.36%, and 89.52%, respectively. We conclude that the newly proposed BASM workﬂow can map burned areas at the subpixel level, providing greater accuracy in regards to the burned area for post-forest ﬁre management and assessment.


Introduction
Forest fires represent one of the biggest threats to plant ecological systems and resources [1], and they are partially responsible for the increase of greenhouse gas in the atmosphere [2,3].Around 350 million hectares of land are affected by forest fires annually due to natural or human causes [4][5][6].Forest fires happen very frequently.A total of 3.52 million hectares of forests were burned in China from 2001 to 2015 according to National Forest Fire Prevention Plan (2016-2025) [7].Mapping the burned area accurately and timely is essential for post-forest fire management and assessment [8], quantifying carbon budgets [9], and for analyzing the relationship between vegetation and climate [9].
The traditional methods for mapping burned area were followed using artificial ground surveys [10] and field sketches [11].These methods for mapping burned area were very inefficient, and these data were unreliable when analyzed at large scale [12].
Satellite-remote sensing can acquire images continuously, even over remote areas, and has the advantage of wide coverage, being timeless, and low cost [13].Burned area mapping based on satellite image has always been a hot spot of research [14][15][16][17].Table 1 summarizes eleven widely used burned-area products.However, the highest spatial resolution of the burned-area products is 30 m, as derived from Landsat data and shown in Table 1.Research on burned-area mapping with high spatial resolution satellite data, such as Sentinel 2 of 10 m spatial resolution, has been actively pursued [6,8,15,29,30].Stroppiana et al. [15] mapped burned areas based on the Sentinel 2 data using an automatic machine learning (ML) algorithm from highly reliable fire points.Florath and Keller [14] detected burned areas using a generic approach based on ML and the Sentinel 2 data.Zhang et al. [6] presented a Siamese self-attention (SSA) classification strategy for burned area mapping using the Sentinel-1 & 2 data.Pinto et al. [8] applied a Burned Areas Neural Network (BA-NET) based method to map burned area using Sentinel-2 and Visible Infrared Imaging Radiometer Suite (VIIRS) data.
For any type of image data, pixel mixing is a common phenomenon [31][32][33].A mixed pixel may contain an unknown composition of land cover types [34], affecting the simulation of radiative characteristics and inversion for land surface parameters from remote sensing data [33].Mixing pixels has been a barrier in the application of remote sensing to burned area mapping.To overcome the barrier, a series of unmixing methods have been proposed, such as fully constrained least squares (FCLS) [35], maximum margin criterion and derivative weights (MDWSU) [36], minimum volume transform (MVT) [37], vertex component analysis (VCA) [38], simplex growing algorithm (SGA) [39], minimum volume constraint nonnegative matrix factorization (MVC-NMF) [40], successive projection algorithm (SPA) [41], etc.He et al. [42] showed that spectral unmixing methods have advantages over the remote sensing vegetation indices due to their ability to decompose the mixed pixel into several fractional abundances.Nascimento and Dias [38] performed a series of experiments using simulated and real data sets and found that the VCA algorithm performs better than the pixel purity index (PPI) method and better than or similar to the N-FINDR algorithms [43].Miao and Qi [40] applied the MVC-NMF method to extract unsupervised endmembers from highly mixed image data.Zhang et al. [41] presented SPA that can provide a general guideline to constrain the total number of endmembers.Kumar et al. [44] estimated global land cover fractions from satellite data based on unconstrained lease squares (UCLS), fully constrained least squares (FCLS), modified fully constrained least squares (MFCLS), simplex projection (SP), sparse unmixing via variable splitting and augmented Lagrangian (SUnSAL), SUnSAL and total variation (SUnSAL TV), and Collaborateve SUnSAL (CL SUnSAL), and their results indicated that FCLS outperformed the other techniques.Thus, in this paper, we used FCLS for spectral unmixing of burned area in mixed pixels.
However, spectral unmixing methods only solve the problem of endmember types and abundances without being able to locate endmembers in a mixed pixel.To solve this problem, Atkinson [45,46] proposed the subpixel concept for precise spatial locations of endmembers in a mixed pixel based on unmixing analysis [47].On the basis of Atkinson's study, many algorithms have been developed [48], such as the spatial attraction model (SAM), double-calculated spatial attraction model (DSAM) [32], subpixel learning algorithms [44], subpixel edge detection method [49], etc.These subpixel techniques have found many applications [50][51][52][53].For instance, the random forests and spatial attraction model (RFSAM) was applied to remote sensing images to improve the accuracy of subpixel mapping of wetland flooding [54].Li et al. [55] predicted a land cover map accurately based on the spatio-temporal subpixel land cover mapping (STSPM) method.Ling et al. [56] monitored variations of reservoir surface water area accurately and timely from daily MODIS images by exploiting information at the subpixel scale.Deng and Zhu [57] successfully applied the Continuous Subpixel Monitoring (CSM) method to map urban impervious surface area (ISA%) at the subpixel scale and characterize its dynamics in Broome County, New York.
However, there are very few applications of these subpixel algorithms for burned area mapping [58].In order to accurately map burned area, the exact location of the burned area in a mixed pixel is important for post-forest fire management and assessment, quantifying carbon budgets, etc.A good subpixel algorithm can help to map and locate burned area with subpixel accuracy, especially from low-spatial-resolution satellite images.However, current burned area mapping algorithms based on satellite remote sensing data can locate burned area only at the pixel level.The objective of this study is to develop a new algorithm to improve burned area mapping at the subpixel level.To this end, we proposed a burned area subpixel mapping (BASM) workflow.The first step is a spectral unmixing analysis using the fully constrained least squares (FCLS) method.The second step is to process the unmixing analysis results with the modified-pixel swapping algorithm (MPSA).The last step is post-classification after MPSA.We then applied the proposed BASM workflow to the Sentinel 2 data sets.The ground-truth fire-scar information provided by the Department of Emergency Management of Hunan Province, China was used as reference data for assessment.

Study Area
Hunan Province, China, is located in the middle Yangtze River, central China, within an area enclosed by 108 • 47 -114 • 15 longitude and 24 • 38 -30 • 08 latitude (Figure 1).Over the course of a year, its temperature typically varies from 7.78 • C to 33.33 • C. The rainiest month is June, with an average rainfall of 177.8 mm.The driest month is December, with an average rainfall of 35.56 mm.The relatively warm and humid weather results in lush vegetation and an evergreen landscape.Hunan Province has a subtropical, evergreen, broad-leaved forest with a forest coverage rate of 59.96%.Hunan Province has a land area of about 211,829 km 2 , and a mountainous area of about 108,472 km 2 accounting for 51.21% of its total.Hunan Province is surrounded by mountains to the east, south, and west sides and a lake basin plain in the north, undulating in the central section, forming an asymmetrical horseshoe shaped terrain opening to the northeast.The east and south sides are mostly at an altitude of above 1000 m.The west side is 500-2099 m above sea level.The central section elevation is below 500 m.Most of the north side elevations are below 50 m.Forest and wetland ecosystems are the two major natural ecosystems.There are about 200 significant ecological zones around the world, and Hunan Province has two forest and wetland ecosystems, i.e., the subtropical, evergreen, broad-leaved forest ecological zones in Wuling Xuefeng Mountains and Nanling Luoxiao Mountains, known as the world's most valuable ecological regions in the same latitude zone.Due to the high forest coverage rate and large mountainous area in Hunan Province, forest fire is a great potential hazard.There are many obstacles for post-forest fire management, e.g., accurate burned area mapping has to rely on manual surveys.To improve this situation, it is imperative to map burned area using automation or semi-automation algorithms, such as remote sensing-based detection and mapping methods.

Methodology
First, we obtained the locations of the 15 fire-scar sites based on the ground-truth fire scar information that was provided by the Department of Emergency Management of Hunan Province.Then, we downloaded the pre-and post-forest fire Sentinel 2 data sets and processed them.Second, burned area fraction was derived from the four selected Multispectral Instrument (MSI)/Sentinel 2 bands listed in Table 2 using the FCLS method [44, 59,60].Taking burned area fraction as input for the Modified Pixel Swapping Algorithm (MPSA) in Visual Studio 2012 platform edited by C# language, we obtained the burned area at subpixel level as output.Furthermore, after a post-classification clump and sieve procedure on the output, the burned area at the subpixel level was mapped.Finally, a comparison analysis and accuracy assessment of the burned area mapped on subpixel level was performed using high spatial resolution GaoFen (GF) series data sets.The flowchart for mapping burned area at the subpixel is shown in Figure 2. Methodological steps are explained in detail over the following subsections.Sentinel 2 is an Earth observation mission of European Space Agency to support environmental services and natural disaster management [61,62].The Sentinel 2 Multispectral Instrument (MSI) data were used in this study, sampling 13 spectral bands: four bands at 10 m, six bands at 20 m, and three bands at 60 m spatial resolutions over the visible, near infrared (NIR), and short wave infrared (SWIR) (https://sentinels.copernicus.eu/web/sentinel/user-guides/sentinel-2-msi/overview, last accessed on 3 April 2022) [62,63].This makes it useful for land monitoring studies at local, regional, and global scales.
The sentinel 2 data were downloaded from United States Geological Survey (USGS, https://earthexplorer.usgs.gov/,last accessed on 1 March 2022) as a Level 1C product, which represent orthorectified, top-of-atmosphere (TOA) reflectance.GF series data sets are important component of China High-Resolution Earth Observation System (CHEOS) project sponsored by China National Space Administration (CNSA) [64].We obtained GF series data sets from the China Centre for Resources Satellite Data and Application (http://www.cresda.com/CN/,last accessed on 2 February 2022) [65].The GF 1 satellite, launched on 26 April 2013, is the first of China's series of high resolution earth observation satellites operated by China [66].The GF 1 satellite is equipped with one four-camera push-broom multispectral system to obtain an 800 km wide image and one two-camera panchromatic/multispectral system (PMS), allowing a temporal resolution of four days [67].Each camera for the four-camera push-broom multispectral system is a wide field of view (WFV) camera with a resolution of 16 m.Thus, the four-camera push-broom multispectral system is also called a wide field of view system (WFVS).The GF 1 PMS has 5 spectral bands, of which one is the panchromatic band (band 1) of 2 m spatial resolution, and the other 4 bands (bands 2-5) are multispectral bands of 8 m resolution, as shown in Table 2.The GF 1B, GF 1C, and GF 1D satellites were launched on 31 March 2018, in coordination with GF 1 to observe earth surface.They have the same sensor systems as GF 1.The PMS on GF 2 also has 5 spectral bands but with higher spatial resolution, e.g., 1 m resolution for the panchromatic band (band 1) and 4 m resolution for the other multispectral bands (bands 2-5), as shown in Table 2.The Sentinel 2, and GF 1, GF 1B, GF 1C, GF 1D, and GF 2 satellite data sets (abbr.GF series data sets) were re-projected to a common Universal Transverse Mercator (UTM) projection with WGS84 as the datum so that all data sets can be overlaid within the same coordinate system.The Sentinel 2 data set has medium spatial resolution and was used as input to our mapping workflow of burned area at the subpixel level, and the GF series data sets have higher spatial resolution and were used as reference data for the accuracy assessment of burned area mapped at subpixel level.The Sentinel 2 satellite data were processed to bottom-of-atmosphere reflectance after radiometric calibration and atmospheric correction using the algorithm 'Sen2Cor' v2.8 (downloadable from http://step.esa.int/main/snap-supported-plugins/sen2cor/sen2cor_v2-8/, last accessed on 3 April 2022) [68][69][70].Table 2 lists the specification of the Sentinel 2 bands (bands 2-4 and 8) selected for this study.

Data Preprocessing
To make full use of spectral information of the selected bands, we carried out two steps further on the Sentinel 2 data sets: (1) Geo-referencing based on Georeferencing tool in ArcMap 10.4 (ESRI) software, adding ten control points to determine the corresponding relationship between the Sentinel 2 data sets and the GF series data sets, and selecting first Order Polynomial (Affine) transformation [71][72][73][74].
To take full advantage of all spectral information, we took four steps to process the GF series data sets (step 1, 3, 4 processed in ENVI 5.3, and step 2 processed in ArcMap 10.4): (1) Geometric orthorectification [77,78] was based on the RPC Orthorectification Batch tool using the Advanced Land-Observing Satellite (ALOS) digital elevation model (DEM) of 12.5 m spatial resolution [79,80].
(2) Geo-referencing was based on the Georeferencing tool by adding ten control points to determine the correspondence between panchromatic band and multispectral bands and selecting first Order Polynomial (Affine) transformation [71][72][73][74].(3) Image fusion was based on the Pan Sharpening Batch tool to fuse panchromatic and multispectral bands to obtain images of 2 m resolution [77,81].
(4) Resampling was based on the GF 2 images at 1 m spatial resolution after image fusion that were resampled to 2 m resolution using the cubic convolution interpolation method [82][83][84].

True Fire-Scar Information
Tabulated fire information collected (TFIC) that was provided by the Department of Emergency Management of Hunan Province was used as ground-truth fire-scar information for this study.Every forest fire event contained information about location, weather, starting and extinguishing times, burned area, and causes, etc.The Department of Emergency Management of Hunan Province confirmed and reported forest fire in details when a forest fire occurred (as shown in Figure 3).Field measurements, including burned area resulted from forest fire, were conducted with Global Navigation Satellite System (GNSS) survey equipment and Unmanned Aerial Vehicle (UAV).The platform was Phantom 4 Pro V2.0, and the sensor was 1-inch CMOS with 20 million effective pixels by well-trained personnel.Well-trained personnel created manually in situ vector data of burned area [85].We transformed the vector data based on the GF series data sets into regions of interest (ROIs) to validate the proposed workflow.Within the TFIC between 2018 and 2021, fifteen fire-scar sites that caused the most environmental damages were selected as reference data for ground truth fire-scar information.Table 3

BASM Approach for Fire Scar Detection
The proposed Burned Area Subpixel Mapping (BASM) workflow consists of three major steps: (1) spectral unmixing using the FCLS Spectral Unmixing tool in ENVI 5.3 software since it outperformed other techniques with the highest classification accuracy and least execution time [44], (2) obtaining burned area spatial distribution information from mixed pixels used modified-pixel swapping algorithm (MPSA), and (3) post-classification after MPSA to denoise the data [86], using the Clump tool in ENVI 5.3 software with Dilate Kernel Value of 3 and Erode Kernel Value of 3. Since burned area mapping at the subpixel level has been rarely investigated, we used the BASM to detect burned area spatial distribution within mixed pixels.The burned area abundance after FCLS was processed with the MPSA to get the final results.Based on the Sentinel 2 satellite data, we divided each pixel into 5 × 5 sub-pixels so that each subpixel has a spatial resolution of 2 m.The final result of burned areas with geographic locations at subpixel level will show the specific spatial distribution of burned areas within a mixed pixel.

Fully Constrained Least Squares (FCLS)
To estimate the abundance of burned area in a mixed pixel in the Sentinel 2 images, we used the well-performed FCLS [35,44,87] method for linear spectral unmixing analysis.There are two constraints imposed on the analysis: (1) endmember abundance sum-to-one ∑ p j=1 α j = 1, and (2) abundance nonnegativity α j ≥ 0 (1 ≤ j ≤ p).We extracted spectra of three variable set of endmembers manually: bare land, forest, and burned area from a Sentinel 2 image, and save them as spectral library (.sli format) for FCLS as input.

Modified Pixel Swapping Algorithm (MPSA)
Atkinson [47] proposed the pixel swapping algorithm (PSA) method that was designed to process an image of land cover proportions in K = 2 classes.In this study, the PSA method was applied to the spectral unmixing results after FCLS.If 10 × 10 (=100) subpixels are to be mapped within each mixed pixel, a land cover class (with or without burned area) with a fraction of 57 percent means that 57 subpixels were allocated to that class.Furthermore, these subpixels are randomly located within the mixed pixel.Once allocated, only the spatial arrangement of the subpixels, not the actual attribute values, can vary.The number of subpixels allocated to that class within each mixed pixel remains fixed.PSA is comprised of three basic steps, but we improved the algorithm as the Modified Pixel Swapping Algorithm (MPSA) that has four steps as detailed below: Firstly, in satellite images with 10 m spatial resolutions, we took burned area as a continuous entity.Within the abundance results after applying the FCLS algorithm, some pixels will appear as false burned area due to noise [60,88,89].In order to minimize misclassification and effects due to noise in the FCLS results on the final burned area at subpixel level, we modified the PSA by adding the object-based approach for classifying burned area, filtered out negative burned area caused by noise, compared with the results from three pixel-based classifiers, and found the most suitable denoising method for this study.We chose a well-known object-based approach and used the three pixel-based classifiers imbedded in the ENVI 5.3 software.Specifically, for the object-based and three pixel-based classifiers, we used four models, including feature extraction rule based (FERB) [90,91], random forest (RF) [92,93], backpropagation neural net (BPNN) [94,95], and support vector machine (SVM) [96,97], that are well-known for their good performance.We manually selected 30% of the clipped image as training samples for RF, BPNN, and SVM, and used the rest as test samples.We took the result of FERB, RF, BPNN, and SVM, respectively, with the result of FCLS as the input of the BASM method.The step of traversing the pixel value of FERB, RF, BPNN, SVM respectively reduced the misclassification and effect due to noise for mapping the burned area at subpixel level.The parameter setup of the four models is shown in Table 4. Secondly, for every subpixel p i , its attractiveness A(p c i ) in the c-th class is predicted as a distance-weighted function of its j = 1,2, . . ., J neighbors: where z(x c j ) is the value of the j-th subpixel belonging to the c-th class (c = 1 or 2), and λ ij is a distance-dependent weight given by the following equation: where h ij is the distance between two subpixels p i and p j , and a is the non-linear parameter of the exponential model.In this study, a is chosen to be 3. Thirdly, once the attractiveness of each subpixel has been calculated based on the current arrangement of subpixel classes, the subpixel algorithm ranks the values on a pixel-by-pixel basis.For each pixel, the least attractive subpixel currently allocated to a "1" (i.e., a "1" surrounded mainly by "0"s) is stored (as shown in Equation ( 4)): The most attractive subpixel currently allocated to a "0" (i.e., a "0" surrounded mainly by "1"s) is also stored (as shown in Equation ( 5)): Lastly, classes of subpixels are swapped as follows: if the attractiveness of the least attractive subpixel is less than that of the most attractive subpixel, the classes are swapped for the subpixels in question to enhance spatial correlation of the sub-pixels (as shown in Equation ( 6)): z( Otherwise no swapping is made.
For the convenience of presenting the results of the proposed BASM traversing through the four models, we named the BASM with FERB as BASM-FERB, with RF as BASM-RF, with BPNN as BASM-BPNN, and with SVM as BASM-SVM, respectively.Furthermore, we directly named BASM without traversing any of the four models as BASM-notra.To demonstrate minimizing misclassification and effect due to noise of the BASM-FERB, BASM-RF, BASM-BPNN, and BASM-SVM, we compared them with BASM-notra.After comparing and analyzing the accuracy assessment of the BASM-FERB, BASM-RF, BASM-BPNN, BASM-SVM, and BASM-notra, the method with the highest accuracy was obtained.

Accuracy Assessment
To evaluate the performance of the BASM workflow, fifteen fire-scar sites with TFIC were selected as reference to assess the accuracy of burned area mapping at subpixel level.For each fire-scar site in a Sentinel 2 image, we used the extracted burned area vector data from the GF series data sets as the reference image due to their high spatial resolution.Five widely used quality indexes, i.e., overall accuracy (OA) [98,99], user's accuracy (UA) [100,101], producer's accuracy (PA) [102,103], intersection over union (IoU) [104,105], and Kappa coefficient (Kappa) [106,107] were calculated for evaluation.OA represents the proportion of the correctly predicted number of pixels to the total.UA represents the proportion of true burned area to the predicted burned area.PA represents the correctly predicted burned area to the true burned area.IoU represents the correct classification accuracy of the models.Kappa represents the agreement between the predictions and ground truth.They can be formulated as follows and calculated from the confusion matrix (as shown in Table 5).Five widely used quality indexes given by the following equation: IoU = TP TP + FP + FN (10) Kappa = OA − Pe 1 − Pe Where, Pe = (TP + FP) × (TP + FN) + (FN + TN) × (FP + TN)
As shown in Figure 4, burned area mapped at the subpixel level based on BASM-FERB (Panel C), BASM-RF (Panel D), BASM-BPNN (Panel E), and BASM-SVM (Panel F) presents less misclassification and effect due to noise than that based on BASM notra (the Panel B).The results indicated that the proposed BASM combining with the four models (classifiers) is effective in minimizing misclassification and effect due to noise.The effect in minimizing misclassification and effect due to noise by BASM-FERB is generally better than BASM-RF, BASM-BPNN, and BASM-SVM.

Pixel and Subpixel Mapping of Burned Area
To illustrate the advantage of the proposed BASM for mapping burned area at subpixel level compared with the traditional object-based method (FERB) and pixel-based method (RF, BPNN, and SVM), we tested it on the fifteen fire-scar sites using the Sentinel 2 satellite image.We compared results from BASM-FERB with FERB, BASM-RF with RF, BASM-BPNN with BPNN, and BASM-SVM with SVM, respectively.We took two fire-scar sites in Hunan Province, China as examples.The first fire-scar site is located in Ningxiang County with 53.06 ha burned area (i.e., 7CSNX) and the second fire-scar site is located in Guiyang County with 50.14 ha burned area (i.e., 8CZGY).
Results for the two examples were shown in Figure 5, in which we compared the burned area mapped at subpixel level with that mapped at pixel level.Comparing panel B with panel G, panel C with panel H, panel D with panel I, and panel E with panel J, we can see that methods FERB, RF, BPNN, and SVM can roughly classify the burned area range (i.e., panels B-E).However, the boundaries of the burned area mapped at pixel level are rough and the visual effect is poor, in contrast to the boundaries of the burned area mapped at subpixel level (i.e., panels G-J) that are clearer and have more spatial details.Results shown in panels B-E, and G-J in Figure 5, indicate that the burned area mapped was not fully classified.Comparing panels B-E, the burned area mapped at pixel level using the FERB (i.e., panel B) is more similar to the reference map (i.e., panel A) than others (i.e., panels C-E).Comparing panels G-J, the burned area mapped at subpixel level using BASM-FERB (i.e., panel G) is more similar to the reference map (i.e., panel F) than others (i.e., panels H-J).(B-E) are the burned area mapped at pixel level, where red area represents burned area but the red area beyond the yellow polygon represents misclassification.Panels (G-J) are burned area mapped at subpixel, in which red area represents burned scar but any red area beyond the yellow polygons represent misclassification.Panel (A) is the fire-scar site based on the GF series data sets.Panels (B-E) are the burned area mapped at pixel level using FERB, RF, BPNN, and SVM, respectively.Panel (F) is the zoomed view of the red circle in Panel (A) based on the GF series data sets.Panels (G-J) are the burned area mapped at subpixel level using BASM-FERB, BASM-RF, BASM-BPNN, and BASM-SVM, respectively.

Accuracy Assessment of the BASM Approach
To further assess the performance of the proposed BASM method, we evaluated it quantitatively at the fifteen fire-scar sites using the Sentinel 2 satellite images.As in Section 4.2, we took two fire-scar sites in Hendong County (i.e., 1HYHD) with 1244.42 ha burned area and Sangzhi County (i.e., 11ZJJSZ) with 34.72 ha burned area.
The results for the Hendong site are shown in Figure 6a.In the panels E-H, the BASM-FERB, BASM-RF, BASM-BPNN, and BASM-SVM have shown classifying ability in the extraction and classification of burned area.Furthermore, as shown in the green circle in panels E-H, the BASM-FERB is better than BASM-RF, BASM-BPNN, and BASM-SVM, which can minimize misclassification in the burned area mapped at subpixel level and can make the burned area mapped at subpixel level accurate.As shown in Table 6, the BASM-FERB outperformed BASM-RF, BASM-BPNN, and BASM-SVM with OA, UA, IoU, and Kappa being 98.56%, 92.70%, 87.56%, and 92.56% respectively, but with PA being 94.04%, slightly lower than that of BASM-RF's, BASM-BPNN's, and BASM-SVM's.

Discussion
At present, the research on burned area mapping is mainly at pixel scale [6,8,29,58,108].Based on the moderate spatial resolution Sentinel 2 data sets, we introduced the concept of subpixel for burned area mapping, and the results had relatively high accuracy.In addition, we modified the pixel swapping algorithm by innovatively traversing the four well-performed classifiers, which reduced the misclassification and effect due to noise of burned area mapped at subpixel level.Traditionally, burned area mapping was done by artificial ground surveys and field sketches [12], which is difficult to cover a large area.Data collection and spatialization is difficult too.Burned area mapping using satellite image data has become a subject of extensive research over the past decades [6,8], and research on using remote sensing techniques to generate products of global burned area started in the late 1980s [109].The advent of remote sensing techniques, either space-borne [15] or airborne [110], gives researchers an alternative or even better way for burned area mapping.
Reduction in noise and misclassification using remote sensing images has always been a classical problem, which has attracted a lot of research interests [111].There are many methods for noise reduction and classification improvement, including second-generation wavelets [112], rank approximation [113], Asymmetrical Gaussian function-fitting [114], Double logistic function-fitting [114], and so on.In the current study, the burned area abundance obtained using FCLS is hindered by ubiquitous noise, thus we modified Pixel Swapping Algorithm by traversing the four well-performed classifiers or models, and then getting the burned area mapped at subpixel level that performed well in noise reduction and classification improvement.
Based on satellite remote sensing image data and techniques, there are many studies using machine learning algorithms for burned area mapping [15,17,29,75,115].Xulu et al. [29] adopt Random Forest classifier to separate burned from unburned area using the Sentinel 2 data.Ramo et al. [17] explored the random forest, support vector machine, neural networks, and decision tree algorithm (C5.0) for classifying burned area at global scale based MODIS data.In addition, researchers used deep learning algorithm [8,30,108,116] to classify burned area from images collected by different sensors.For example, Seydi et al. [30] proposed a novel framework based on deep Siamese morphological neural network (DSMNN-Net) for burned area mapping.Zhang et al. [6] presented a Siamese self-attention classification strategy for the multi-sensor burned area mapping.However, whether using machine learning algorithms or deep learning algorithms based on remote sensing data for burned area mapping, most current research basically performs the mapping at pixel level, and very rarely at subpixel level.In this paper, we proposed the BASM workflow and evaluated it at fifteen fire-scar sites, and the results indicated that using BASM-FERB, BASM-RF, BASM-BPNN, and BPSM-SVM for mapping burned area at subpixel level provides finer resolution when compared to FERB, RF, BPNN, and SVM at pixel level.The efficiency of the BASM-FERB, BASM-RF, BASM-BPNN, and BPSM-SVM was roughly the same, and each method took less than 30 min from the start to the output of results.In addition, burned area mapped at the subpixel level using BASM-FERB agreed very well to the reference data.
Furthermore, burned area mapping at pixel level could cause large errors.Ability to specify the distribution of endmember composition in a mixed pixel can improve such a situation [46].Ruescas et al. [117] estimated burnt land percentage at subpixel level using Advanced Very High Resolution Radiometer (AVHRR) data with a mean error of 6.5%.In this paper, we selected endmembers manually, which was not as efficient as automatic or semiautomatic procedures and was subject to operator experience.It needs to be considered to adopt automatic or semiautomatic procedures for endmember selection, such as the Pixel Purity Index (PPI) and Sequential Maximum Angle Convex Cone (SMACC) algorithms in ENVI.Furthermore, we derived burned area abundance from the Sentinel 2 data and further output the specific spatial distribution of burned area in mixed pixels.It is widely known that, in mountainous areas, burned area mapping based on unmixing models is confounded by shadows.For example, as shown in Figure 4, the misclassification and effect due to noise, including any red area beyond the yellow line, may be caused by shadows.In order to minimize misclassification and effects confounded by shadows, we modified the PSA by traversing the result of four approaches respectively, then filtered out the negative burned area caused by noise, such as shadows.In this study, we proposed the BASM workflow for mapping burned area at the subpixel level and we compared BASM-FERB, BASM-RF, BASM-BPNN, BASM-SVM, and BASM-notra with each other.Except for the case that the BASM-FERB's average PA (=89.52%) is slightly lower than that of BASM-RF (average PA = 89.97%)and BASM-notra (average PA = 91.36%),BASM-FERB's OA, UA, IoU, and Kappa are significantly higher than those of the other four methods.It is of great significance for post-forest fire management and assessment to derive burned area at the subpixel level from remote sensing data.Moreover, the BASM-FERB method was successfully applied to the fifteen fire-scar sites in Hunan Province, China.

Conclusions
We proposed an approach named BASM to map burned area at the subpixel level using Sentinel 2 image data.The proposed method contains three main steps: (1) spectral unmixing used FCLS, (2) derivation of burned area at subpixel level using MPSA, and (3) post-classification after MPSA using Clump.The BASM was applied to fifteen fire-scar sites located in Hunan Province, China.Comparing with the traditional mapping methods (FERB, RF, BPNN, and SVM) for burned area mapping at pixel level, BASM-FERB, BASM-RF, BASM-BPNN, and BASM-SVM generated burned area as mapped at the subpixel level.In terms of minimizing misclassification and effects due to noise, a comparison between BASM-notra, BASM-FERB, BASM-RF, BASM-BPNN, and BASM-SVM could optimize the selection of methods to reduce noise effectively.Results showed that BASM-FERB has better mapping accuracy of burned area at the subpixel level, which is reflected in four accuracy evaluation indices i.e., OA, UA, IoU, and Kappa.The mapping accuracy of burned area at subpixel level based on the BASM-FERB has been largely enhanced with average OA (98.11%), average UA (81.72%), average IoU (74.32%), and average Kappa (83.98%), respectively.Even though BASM-RF's and BASM-notra's average PA (89.97%), (91.36%) is higher than BASM-FERB's (89.52%), the difference is very small.We concluded that: (1) the BASM method developed in this study has well demonstrated its better performance in mapping burned area at the subpixel level compared to conventional methods at the pixel level, (2) the BASM in combination with any of the four well-performing classifiers can reduce misclassification and minimize noise effectively when compared with the results obtained by BASM alone, and (3) the BASM-FERB can generate more accurate burned area as mapped at subpixel level than BASM-RF, BASM-BPNN, BASM-SVM, and BASMnotra.This new mapping technique for mapping burned area at the subpixel level can enhance post-fire management and assessment and help in quantifying carbon budgets.
Note: No. = number.ESA = European Space Agency.MODIS = Moderate-resolution Imaging Spectroradiometer.AVHRR = advanced very high resolution radiometer.PROBA = Project for On-Board Autonomy.MCD64A1 = MODIS Direct Broadcast Monthly Burned Area Product Collection 6. NASA = National Aeronautics and Space Administration.GWIS = Global Wildfire Information System.JRC = Joint Research Centre.GABAM = Global Annual Burned Area Mapping.IRSDE/CAS = Institute of Remote Sensing and Digital Earth-Chinese Academy of Sciences.OLI = Operational Land Imager.TM = Thematic mapper.ETM = Enhanced Thematic Mapper.TREES = Tropical Ecosystems and Environmental Sciences lab. INPE = National Institute for Space Research.DETER = near real-time deforestation detection.AWiFS = Advanced Wide-Field Sensor.

Figure 1 .
Figure 1.Study site with locations of the fifteen fire-scar sites in Hunan Province, China.

Figure 2 .
Figure 2. Flowchart for the analysis performed in this study.

Figure 3 .
Figure 3. Field measurement confirmation and reports in details by well-trained personnel.Panels (A,B) show the fire scar sites.Panel (B) show that the trained personnel were using UAV (the UAV is shown in red circle) and other equipment for field surveys.Panel (C) shows the manually created vector data of burned area.

Figure 4 .
Figure 4. Comparison of BASM-notra with BASM-FERB, BASM-RF, BASM-BPNN, and BASM-SVM in minimizing misclassification and effect due to noise for Sentinel 2 image at eleven fire scar sites.The yellow line of panels (A-F) are fire-scar sites vector data.Panels (B-F) are results of burned area mapped at subpixel level, in which red represents burned area.Any red area beyond the yellow line represents misclassification and effect due to noise.Panel (A) is the fire-scar site based on the GF series data sets.Panel (B) is based on BASM-notra.Panel (C) is based on BASM-FERB.Panel (D) is based on BASM-RF.Panel (E) is based on BASM-BPNN.Panel (F) is based on BASM-SVM.

Figure 5 .
Figure 5.Comparison of the burned area mapped at subpixel level with that mapped at pixel level using different algorithms at the fire-scar site in 7CSNX (a), 8CZGY (b).The yellow polygon of panels (A-J) are the fire-scar site vector data.Panels (B-J) are zoomed view of the red circle in Panel (A).Panels (B-E) are the burned area mapped at pixel level, where red area represents burned area but

Figure 6 .
Figure 6.The burned area mapped at subpixel level using different algorithms in the Sentinel-2 image over fire-scar site in 1HYHD (a), 11ZJJSZ (b).The yellow polygon in panels (A-H) is the vector data of fire-scar sites.Panels (B-H) are zoomed view of the red circle in Panel (A).Panels (E-H) are burned area mapped at subpixel level, in which the red area represents burned area and purple color represents background.Panel (A) shows the fire-scar site derived from the GF series data sets.Panel (B) shows the pre-fire site based on the Sentinel-2 images.Panel (C) shows the post-fire site.Panel (D) is the post-fire site based on the GF series data sets.Panel (E) is the burned area mapped at subpixel using BASM-FERB.Panel (F) shows the burned area mapped at subpixel level using BASM-RF.Panel (G) is the burned area mapped at subpixel level using BASM-BPNN.Panel (H) is the burned area mapped at subpixel used BASM-SVM.

Table 1 .
Overview of eleven widely used burned-area products.

Table 2 .
Selected bands of Sentinel 2, GF 1, GF 1B, GF 1C, GF 1D, and GF 2 satellite imagery data sets used in the current study.

Table 3 .
Fire information at the fifteen selected sites.

Table 5 .
Confusion matrix of ground truth and prediction.