An automatic procedure for generating burn severity maps from the satellite images-derived spectral indices

Full Terms & Conditions of access and use can be found at https://www.tandfonline.com/action/journalInformation?journalCode=tjde20 International Journal of Digital Earth ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tjde20 An automatic procedure for generating burn severity maps from the satellite images-derived spectral indices Saeid Gholinejad & Elahe Khesali To cite this article: Saeid Gholinejad & Elahe Khesali (2021): An automatic procedure for generating burn severity maps from the satellite images-derived spectral indices, International Journal of Digital Earth, DOI: 10.1080/17538947.2021.1966525 To link to this article: https://doi.org/10.1080/17538947.2021.1966525 Published online: 18 Aug 2021. Submit your article to this journal View related articles View Crossmark data An automatic procedure for generating burn severity maps from the satellite images-derived spectral indices Saeid Gholinejad a and Elahe Khesali b a Department of Geomatics Engineering, Faculty of Civil Engineering and Transportation, University of Isfahan, Isfahan, Iran; b Remote Sensing & Photogrametry Department, Faculty of Geodesy and Geomatics Engineering, K.N. Toosi University of Technology, Tehran, Iran ABSTRACT Fire, especially wildfire, which can be considered as one of the main threats to vegetation cover and animals’life, has attracted lots of attention from environmental researchers. To better manage the fire crisis and take the necessary measures to compensate for its damages, it is essential to have detailed information about the burn severity levels. Accordingly, satellite images and their spectral indices have been widely considered in the literature as powerful tools in producing burn severity information. Despite the efficiency of the previously proposed methods, the necessity of ground reference data for their thresholding step faces them with serious challenges. To address this problem, in this study, an automatic procedure based on the change-point analysis is presented for thresholding differenced normalized burn ratio (dNBR) and its another version, dNBR2. In this procedure, a mean-shift based change-point analysis is performed on the dNBR and dNBR2 images for classifying them into burn severity levels. Experiments, conducted on some parts of Alaska and California in the United States, illustrated the high efficiency of the proposed method. Moreover, as an applied experiment, the severity of the fires, occurred in 2020 in the Khaeiz protected area in Iran, was estimated and compared with local reports. ARTICLE HISTORY Received 6 February 2021 Accepted 5 August 2021 KEYWORDS Burn severity mapping; spectral indices; changepoint analysis; normalized burn ratio (NBR) 1. Introduction Every year, millions of hectares of forests and vegetation areas around the world are destroyed by fire, which has a significant impact on the ecosystem and habitat of animals in these areas. These fires often occur in the warm seasons of the year, mainly due to climate change, human factors, and special weather conditions such as drought, strong winds, and heat waves (FernandezManso, Quintano, and Roberts 2019). In order to compensate for the loss to the vegetation after the fire, it is very important to determine the location and the severity of the fires. To this end, field-based surveying is the initial strategy. But unlike the high accuracy of this strategy, it is very costly and time-consuming and is not proper for real-time applications. Due to these problems, in recent decades, satellite imagery with sampling from different part of the electromagnetic spectrum, have a lower cost than the field-based method, covering the wide spatial extent, and have time-series images from different regions have been introduced as a suitable alternative to fieldbased surveying (Quintano, Fernandez-Manso, and Roberts 2017). Accordingly, these images have been significantly used in the literature to detect burn severity. In general, models for burn © 2021 Informa UK Limited, trading as Taylor & Francis Group. The International Journal of Digital Earth is an Official Journal of the International Society for Digital Earth CONTACT Saeid Gholinejad [email protected] Department of Geomatics Engineering, Faculty of Civil Engineering and Transportation, University of Isfahan, Isfahan 8174673441, Iran INTERNATIONAL JOURNAL OF DIGITAL EARTH https://doi.org/10.1080/17538947.2021.1966525 severity mapping can be categorized into two main categories: physical models and empirical models (He et al. 2019). Physical models, like radiative transfer models, simulate the physical interaction between the canopies of the burned vegetation and light radiation (Chuvieco et al. 2006). These models have a complex implementation and, as their name implies, require some physical parameters of the scene of imagery. In contrast, empirical models, like spectral indices-based models, are easy to implement and interpret. Due to this simplicity, these models, especially spectral indices-based ones, have been highly considered in a vast extent of studies to detect the burn severity of the burnt areas. Several spectral indices have been used to detect burned area from satellite images. Burned area index (BAI) (Chuvieco, Martin, and Palacios 2002), mid infra-red burn index (MIRBI) (Trigg and Flasse 2001), normalized burned ratio (NBR) (Key and Benson 2006), normalized burned ration 2 (NBR2) (Lutes et al. 2006), normalized burn ratio thermal (NBRT1) (Holden et al. 2005) are the most important indices, used for burn detection in the literature. Among these different spectral indices, differenced version of NBR (dNBR) has attracted lots of attention in burn severity level mapping, so that it is known as a reference methodology (Soverel, Perrakis, and Coops 2010). Along with dNBR, relative dNBR (RdNBR) (Miller and Thode 2007) is another commonly used method for burn severity estimation in different studies. A multi-temporal dNBR method was proposed in Veraverbeke et al. (2011) based on the MODIS images for the large 2007 Peloponnese Region wildfire in Greece. Chu et al. also proposed a procedure based on the differencing of vegetation indices including dNBR, differenced normalized difference vegetation index (dNDVI), and differenced normalized difference moisture index (dNDMI) to determine the burn severity in the Siberian boreal larch forest in northern Mongolia within different time lags (Chu, Guo, and Takeda 2016). Furthermore, Quintano et al. applied dNBR and RdNBR, along with relative burn ratio (RBR) (Parks, Dillon, and Miller 2014), to map burn severity in a region in Spain using the combination of Landsat-8 and Sentinel-2 data sets (Quintano, Fernández-Manso, and Fernández-Manso 2018). In another study, burned areas and their corresponding burn severity levels of the sugarcane area were extracted using dNBR images in a region in Tarlac, Philippines (Baloloy et al. 2016). In 2020, Saulino used Landsat-8 and Sentinel-2 derived dNBR and investigated their performance using field-based spectral indices (Saulino et al. 2020). They showed the superiority of Sentinel-2 derived dNBR over Landsat-8 one for burn severity mapping in the Mediterranean wildfire. As can be inferred from the above studies, dNBR is an important tool for burn severity map generation due to its simplicity and effectiveness. However, to achieve a burn severity map from a dNBR image, some threshold values should be applied to distinguish the different levels of burn severity, which is a challenging procedure. It is noteworthy that the greater the dNBR value is, the greater the level of the burn will be. Accordingly, different thresholding procedures were used in burn monitoring studies. Key and Benson proposed burn severity ranges based on a field-based matching as Table 1 (Key and Benson 2006). Such a classification has been used in different studies. However, these threshold Table 1. Threshold values provided by Key and Benson for burn severity determination through dNBR image (Key and Benson 2006). Burn severity level dNBR range Enhanced regrowth, high −0.5 to −0.251 Enhanced regrowth, low −0.25 to −0.101 Unburned −0.1 to 0.099 Low severity 0.1 to 0.269 Moderate-low severity 0.27 to 0.439 Moderate-high severity 0.44 to 0.659 High severity 0.66 to 1.3 2S. GHOLINEJAD AND E. KHESALI values are case dependent, and they are not optimal for different scenarios. Like this study, a nonlinear regression model between dNBR and field-based burn indices, including composite burn index (CBI) (Key and Benson 2006) and its geometrically improved version (GeoCBI) (De Santis and Chuvieco 2009), was proposed for the Mediterranean forest type in Saulino et al. (2020). Boucher et al. successfully used a polynomial model between the field data and spectral indices to determine threshold values (Boucher et al. 2017). An iterative procedure was also used in Miller and Thode (2007) to classify burn severity levels, in which increasing values for each threshold were combined until the highest proportion of samples was correctly classified. The last two procedure also were used in María Guadalupe, Mundo, and Veblen (2020) for burn severity monitoring in North Patagonian Forests. Although the above-mentioned procedure can to some extent properly classify the different levels of burn severity, it requires some training field-based data. As the study area expands, it will be practically impossible to obtain terrestrial data for finding the optimal threshold values. On the other hand, there is no information about the reliability and sufficiency of the field samples. To address the challenging problem of automatically determining proper threshold values, without field-based samples, a burn severity mapping procedure has been proposed in this study based on the satellite-derived dNBR and dNBR2 images and change-point analysis (Eckley, Fearnhead, and Killick 2011). It should be noted that the main goal of this study is to demonstrate the potential of change-point analysis for the classification of the difference bands of spectral indices into burn severity levels. To do so, we used two spectral indices, NBR and NBR2, to show the performance of the proposed thresholding process. In the proposed procedure, after masking vegetational increased regions and water bodies, NBR and NBR2 bands for preand post-fire images were calculated. Then, the dNBR and dNBR2 bands were generated by differencing the mentioned spectral indices. After that, the change-point analysis method was performed on the dNBR and dNBR2 bands to find three optimal points, in which the mean statistic of the data changed. These three points divided the data range of dNBR and dNBR2 into four intervals: unburned, low severity, medium severity, and high severity. The remainder of this study is organized as follows. In Section 2, we describe the detailed information of our proposed burn severity mapping procedure. Study areas, data sets, and the results of implementing our proposed procedure are explained in Section 3. The conclusion and future remarks are presented in Section 4. 2. Methodology The proposed procedure for burn severity mapping consists of two main parts: (1) generate difference bands of spectral indices, and (2) thresholding difference bands using the change-point analysis to classify them into burn severity levels. Figure 1 shows the general workflow of the burn severity mapping in this study. In the following, the detailed steps of the main two parts of the proposed procedure are introduced. 2.1. Generating difference bands of burn indices As shown in Figure 1, the first step of the proposed method is generating mosaic preand post-fire images. This step is performed due to working with a large study area and probable lack of images at a specified required time. To make sure the produced mosaic images (i.e. preand post-fire images) cover the entire study area, two-month period images were collected corresponding to preand post-fire. The second step is removing vegetational-increased regions from the images that implies there is no fire in these areas. To this end, dNDVI between preand post-fire mosaic images were produced INTERNATIONAL JOURNAL OF DIGITAL EARTH 3 as follows: NDVI = r NIR − r R r NIR + r R dNDVI =NDVIpost−fire −NDVIpre−fire (1) Figure 1. Flowchart of the proposed procedure for burn severity mapping. 4S. GHOLINEJAD AND E. KHESALI where r NIR and r Rare respectively the reflectance of the near-infrared and red bands. Then, for ith pixel of the image: dNDVIi.0, vegetation increased pixel dNDVIi,0, otherwise. (2) Water bodies are masked from the preand post-fire images in the third step of the proposed procedure. In this step, the normalized difference water index (NDWI) for mentioned images calculated as follows (McFeeters 2013): NDWI = r G− r NIR r G+ r NIR (3) where r Gis the reflectance in the green band of the image. As stated by McFeeters (2013), pixels with NDWI values more than 0.3 are considered as the water bodies. Our experiments proved that this threshold value is the optimal one for our case studies. The next step of our proposed burn severity mapping procedure is calculating burn index bands, which are NBR and NBR2 in our study. These two bands are produced by using the following equations (Key and Benson 2006; Lutes et al. 2006). NBR = r NIR − r SWIR2 r NIR + r SWIR2 (4) and NBR2= r SWIR1− r SWIR2 r SWIR1+ r SWIR2 (5) where r SWIR1and r SWIR2are the reflectance of the short wave infrared bands, i.e. SWIR1 and SWIR2. It is worth noting that all the satellite images do not have these two bands. Consequently, Equation (5) is used for situations where the satellite has both SIWR bands, such as the Landsat-8 satellite images. Finally, the differences between preand post-fire burn indices bands were calculated as follow: dNBR =NBRpre−fire −NBRpost−fire dNBR2=NBR2pre−fire −NBR2post−fire (6) 2.2. Change-point analysis based thresholding Change points analysis is the procedure of extracting some points from a data series, in which the distributional properties of the data series changed. Therefore, this analysis can divide the data series into some statistically homogenous regions (Eckley, Fearnhead, and Killick 2011). Basically, change-point analysis is for time-series data, in which a set of data is monitored over time. In this study, this analysis was applied in the spatial domain. A mean-shift based change-point analysis model (Page 1955) was performed for burn severity mapping purpose. Let x1,...,xnare increasingly ordered sequence of pixel values of the difference images, i.e. dNBR and dNBR2. The values of the mentioned sequence can be written in the following form: xi= m + d i+ e i;i=1, ...,n(7) where μis the mean of the sequence, before the unknown change-point occurred in the location c. e iis the zero-mean independent random error with unknown variance in the INTERNATIONAL JOURNAL OF DIGITAL EARTH 5 location i. Moreover, d iis defined as follows: d i=01≤i≤c Dc,i≤n (8) where nis the total number of gray levels in the image. To detect mean shifts, cumulative sum (CUSUM) is the first strategy which is calculated as follows (Page 1955): CUSUM(m)=1 n √ m i=1 xi−m n n i=1 xi  m=1, ...,n(9) where mis a gray level of the image pixels. The expectation of the CUSUM(m)isasfollows: E[CUSUM(m)] =n−3/2m(n−m)D(10) From the above equation, it is inferred that if |CUSUM(m)|is large, the Δvalue is non-zero, and therefore, a change-point has occurred in the location m.Asaresult,itcanbesaidthat thechangepointisthepoint,inwhichtheCUSUM statistic is maximized. Multiple change points are determined using the points with higher CUSUM statistics. In this study, three change-points, c1,c2, and c3, are extracted as threshold values for burn severity map generation. Based on these threshold values, the burn severity of the jth pixel of the difference image (dNBR or dNBR2) is categorized as follows: Burn severityj= Unburned DIj≤c1 Low severity c1,DIj≤c2 Moderate severity c2,DIj≤c3 High severity c3,DIj ⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩ (11) 3. Experiments and analysis In this section, to evaluate the performance of the proposed procedure, different experiments have been conducted in two main parts. In the first part of the experiments, the accuracy of the results of the proposed algorithm has been analyzed using reference burn severity maps, generated under the monitoring trends in burn severity (MTBS) project (Eidenshink et al. 2007). After revealing the effectiveness of the proposed method for burn severity mapping in the first part, in the second part of the experiments, the burn severity map for the Khaeiz protected area in Iran, which caught fire in the spring of 2020, has been produced and compared to the local reports. 3.1. Experiments using ground truth Regions in Alaska and California in the United States have been used to evaluate the effectiveness of the proposed burn severity mapping method. In these two regions, several fires have been occurred in 2018 that caused a large amount of vegetation destruction. Landsat-8 satellite images have been used to determine the burn severity in the study regions. To form preand post-fire images, we collected all the images with less than %20 cloud coverage, acquired in September and October from the two mentioned regions. September and October are the last months of the warm season of the year, and as a result, the damages of fires of the warm seasons can be monitored in their corresponding images. On the other hand, in these months, the amount of clouds and snow in the images is low enough to efficiently analyze fire-based changes on the earth’s surface. After collecting mentioned images, we generated preand post-fire mosaic images for the years 2017 and 2018, which reflect the severity of the 2018 burnings. Then, the proposed algorithm was performed on the mosaic images to obtain burn severity maps. 6S. GHOLINEJAD AND E. KHESALI The data, provided under the MTBS project, were used as reference data in this study. This data, which is freely available to users (https://www.mtbs.gov/direct-download), is based on the NBR index and the interpretation of its values, as well as using CBI ground data. The burn severity maps in the MTBS project consist of six classes including unburned, low severity, moderate severity, high severity, increased greenness, and water bodies. In our experiments, we discarded two increased greenness and water bodies classes, because our proposed burn severity mapping method already masked these areas during its procedures. To provide a competitive qualification for the proposed method, we used the traditional dNBR thresholding method (Key and Benson 2006), which is the only available unsupervised option that has a similar purpose to our research. This method simply segments dNBR band with some of the global suggested threshold values as described in Table 1. Figure 2 shows the map of the severity of the burn, occurred in 2018 in a region in Alaska. As shown in this figure, there is a high similarity between the obtained maps through the proposed method and the reference map, especially in areas with high burn severity. Since there is no information about the exact study period of the reference burn severity map, the obtained maps and the reference map do not exactly match each other in terms of the study period. Hence the slight Figure 2. Burn severity maps obtained for a region in Alaska through (a) our proposed method by dNBR, (b) our proposed method by dNBR2, (c) traditional dNBR classification proposed by Key and Benson (2006), and (d) the reference burn severity map provided by MTBS project. INTERNATIONAL JOURNAL OF DIGITAL EARTH 7 differences between these maps are inevitable. Although the dNBR and dNBR2 images were not similar, and also, their threshold values were not the same, however, their obtained burn severity maps are very similar. This indicates the high performance of the proposed methods, regardless of the applied spectral index. The visual interpretation of the maps in Figure 2 also shows that the similarity of the dNBR-based burn severity map to the reference map was higher than the dNBR2-based one. This higher similarity can be seen by comparing the moderate burn severity parts of these maps with the reference map. To more detailed examine the differences between the generated maps, three sub-scenes from the Alaska region were selected and the results are shown separately in Figure 3. As shown in this figure, there are considerable differences between the results of the proposed method with those of traditional dNBR method. In these sub-scenes, the traditional dNBR method places most of the pixels at the moderate severity level. For example, for the first sub-scene (first row), the traditional dNBR method placed a large number of pixels with low burn severity at the moderate burn level. In the second sub-scene, regions with high burn severity located at the moderate level, while the proposed method accurately detected the severity of burns in these areas. This is also the case for the third sub-scene, and unlike the traditional dNBR method, the proposed method had a high accuracy in identifying areas with high burn severity. Tables 2 and 3show the confusion matrices of burn severity mapping obtained by our proposed method on the dNBR and dNBR2 bands, respectively. As shown in the tables, the highest user’s accuracy belonged to the high severity class. Moreover, in both cases, the lowest accuracy values obtained for low severity classes. In practice, there is not much difference between low burn severity and unburned classes. The most important aspect of the burn severity mapping is to distinguish between pixels of low burn severity (or unburned) class and those of high burn severity class. Figure 3. Burn severity maps for three sub-scenes of the Alaska region. Upper figure shows the location of the sub-scenes. Rows from the top to the bottom are first to third sub-scenes, and columns from left to right corresponds to: dNBR-based proposed method, dNBR2-based proposed method, traditional dNBR, and reference map. 8S. GHOLINEJAD AND E. KHESALI Eidenshink, Jeff, Brian Schwind, Ken Brewer, Zhi-Liang Zhu, Brad Quayle, and Stephen Howard. 2007.“A Project for Monitoring Trends in Burn Severity.”Fire Ecology 3 (1): 3–21. Fernandez-Manso, Alfonso, Carmen Quintano, and Dar A. Roberts. 2019.“Burn Severity Analysis in Mediterranean Forests Using Maximum Entropy Model Trained with EO-1 Hyperion and LiDAR Data.”ISPRS Journal of Photogrammetry and Remote Sensing 155: 102–118. He, Yinan, Gang Chen, Angela De Santis, Dar A. Roberts, Yuyu Zhou, and Ross K. Meentemeyer. 2019.“A Disturbance Weighting Analysis Model (DWAM) for Mapping Wildfire Burn Severity in the Presence of Forest Disease.”Remote Sensing of Environment 221: 108–121. Holden, Z. A., A. M. S. Smith, P. Morgan, M. G. Rollins, and P. E. Gessler. 2005.“Evaluation of Novel Thermally Enhanced Spectral Indices for Mapping Fire Perimeters and Comparisons with Fire Atlas Data.”International Journal of Remote Sensing 26 (21): 4801–4808. Key, Carl H., and Nathan C. Benson. 2006.“Landscape Assessment (LA).”In FIREMON: Fire Effects Monitoring and Inventory System. Gen. Tech. Rep. RMRS-GTR-164-CD, Fort Collins, CO: US Department of Agriculture, Forest Service, Rocky Mountain Research Station. p. LA-1-55, edited by Duncan C. Lutes, Robert E. Keane, John F. Caratti, Carl H. Key, Nathan C. Benson, Steve Sutherland, Larry J. Gangi, 164. Lutes, Duncan C., Robert E. Keane, John F. Caratti, Carl H. Key, Nathan C. Benson, Steve Sutherland, and Larry J. Gangi. 2006.“FIREMON: Fire Effects Monitoring and Inventory System.”In Gen. Tech. Rep. RMRS-GTR-164. Fort Collins, CO: US Department of Agriculture, Forest Service, Rocky Mountain Research Station. 1 CD., 164. María Guadalupe, Franco, Ignacio A. Mundo, and Thomas T. Veblen. 2020.“Field-Validated Burn-Severity Mapping in North Patagonian Forests.”Remote Sensing 12 (2): 214. McFeeters, Stuart K. 2013.“Using the Normalized Difference Water Index (NDWI) Within a Geographic Information System to Detect Swimming Pools for Mosquito Abatement: A Practical Approach.”Remote Sensing 5 (7): 3544–3561. Miller, Jay D., and Andrea E. Thode. 2007.“Quantifying Burn Severity in A Heterogeneous Landscape with a Relative Version of the Delta Normalized Burn Ratio (dNBR).”Remote Sensing of Environment 109 (1): 66–80. Page, E. S. 1955.“A Test for A Change in A Parameter Occurring At An Unknown Point.”Biometrika 42 (3/4): 523– 527. Parks, Sean A., Gregory K. Dillon, and Carol Miller. 2014.“A New Metric for Quantifying Burn Severity: The Relativized Burn Ratio.”Remote Sensing 6 (3): 1827–1844. Quintano, C., Alfonso Fernández-Manso, and Oscar Fernández-Manso. 2018.“Combination of Landsat and Sentinel-2 MSI Data for Initial Assessing of Burn Severity.”International Journal of Applied Earth Observation and Geoinformation 64: 221–225. Quintano, Carmen, Alfonso Fernandez-Manso, and Dar A. Roberts. 2017.“Burn Severity Mapping From Landsat MESMA Fraction Images and Land Surface Temperature.”Remote Sensing of Environment 190: 83–95. Saulino, Luigi, Angelo Rita, Antonello Migliozzi, Carmine Maffei, Emilia Allevato, Antonio Pietro Garonna, and Antonio Saracino. 2020.“Detecting Burn Severity Across Mediterranean Forest Types by Coupling MediumSpatial Resolution Satellite Imagery and Field Data.”Remote Sensing 12 (4): 741. Soverel, Nicholas O., Daniel D. B. Perrakis, and Nicholas C. Coops. 2010.“Estimating Burn Severity From Landsat DNBR and RdNBR Indices Across Western Canada.”Remote Sensing of Environment 114 (9): 1896–1909. Trigg, S., and S. Flasse. 2001.“An Evaluation of Different Bi-Spectral Spaces for Discriminating Burned ShrubSavannah.”International Journal of Remote Sensing 22 (13): 2641–2647. Veraverbeke, Sander, Stefaan Lhermitte, Willem W. Verstraeten, and Rudi Goossens. 2011.“A Time-Integrated MODIS Burn Severity Assessment Using the Multi-Temporal Differenced Normalized Burn Ratio (dNBRMT).”International Journal of Applied Earth Observation and Geoinformation 13 (1): 52–58. INTERNATIONAL JOURNAL OF DIGITAL EARTH 15