Remote Sensing of Forest Burnt Area, Burn Severity, and Post-Fire Recovery: A Review

: Wildland ﬁres dramatically affect forest ecosystems, altering the loss of their biodiversity and their sustainability. In addition, they have a strong impact on the global carbon balance and, ultimately, on climate change. This review attempts to provide a comprehensive meta-analysis of studies on remotely sensed methods and data used for estimation of forest burnt area, burn severity, post-ﬁre effects, and forest recovery patterns at the global level by using the PRISMA framework. In the study, we discuss the results of the analysis based on 329 selected papers on the main aspects of the study area published in 48 journals within the past two decades (2000–2020). In the ﬁrst part of this review, we analyse characteristics of the papers, including journals, spatial extent, geographic distribution, types of remote sensing sensors, ecological zoning, tree species, spectral indices, and accuracy metrics used in the studies. The second part of this review discusses the main tendencies, challenges, and increasing added value of different remote sensing techniques in forest burnt area, burn severity, and post-ﬁre recovery assessments. Finally, it identiﬁes potential opportunities for future research with the use of the new generation of remote sensing systems, classiﬁcation and cloud performing techniques, and emerging processes platforms for regional and large-scale applications in the ﬁeld of study.


Introduction
Wildland fires are periodic disturbance events that dramatically affect the structure and distribution of global forest ecosystems, altering soil erosion, causing loss of biodiversity, habitat, production and productivity, endangering human life, and disrupting livelihoods [1].Over the past two decades, 7.20 billion ha of land was burned globally by wildland fires at an average of 400 million ha annually [2].According to an analysis by the Food and Agricultural Organization, between 2001 and 2019, forest (tree-covered) areas accounted for about 29% of the total area burnt by wildfires [3].The largest burnt treecovered area in the total wildland fire area was in Central America (47%), while the smallest was in the Oceania region (2%) [3].
Furthermore, wildland fires have a strong influence on climate change due to the rapid release of carbon sequestered in forest biomass into the atmosphere [4,5].In recent decades, the rapidly changing climate has increased the frequency and severity of wildland fires [6][7][8][9][10], and it is projected that further warming effects will likely cause more dramatic consequences to the forests, and an increase of burnt area (BA) [11][12][13][14][15].
Growing concerns about the impact of climate change has increased the need for the development of sustainable forest management policies aiming to mitigate greenhouse gas emissions and increase the efficiency of fire management [16][17][18].Broad scale severity levels at which fire affects a forest cover is a key factor for the successful assessment of postfire impact and the strategy that promotes natural regeneration, as well as for understanding the effects of climate change on ecosystems [19,20].Thus, comprehensive burnt area, burn severity (BS), and post-fire recovery monitoring and assessments are a critical tool for researchers, foresters, and fire managers.
Over the last few decades, the use of space-borne sensor data has provided unique advances for the professional community to detect forest ecosystem disturbances, especially for estimations of spatio-temporal fire dynamics and BA mapping [21,22].Remote sensing (RS) techniques become an unprecedented alternative option to expensive and time-consuming field measures in the monitoring of forest ecosystems over large and remote geographic areas [23][24][25][26].Multi-temporal satellite images of broad spatial resolution have been extensively used in measurements of forest areas prior to and post-fire acquisitions to map BA at regional and global scales [27][28][29][30], and to estimate burn severity [31][32][33][34] and vegetation recovery [35][36][37][38].
Given its importance for monitoring forest ecosystems and environmental management policies, the main objective of this paper is a review of state-of-the-art remotely sensed methods and data for estimation of forest BA, BS, and forest recovery patterns at the global level that have been applied in the research papers published in English in peerreviewed scientific journals within the period between 2000 and 2020.The paper analyses current research trends and summarizes the application of remote sensing techniques for the burnt forest area, burned severity, and post-fire recovery.Finally, we recommend future research for the use of the new generation of RS systems, classification, and cloud performing techniques, and emerging process platforms for the regional and large-scale applications in BA studies.

Search Methodology
A systematic approach of a literature review from 2000 to 2021 was used in this research with a focus only on scientific peer-reviewed papers in the fields of BA, remote sensing, and forest ecosystems at the local, regional, and global levels.We selected this timeframe for two reasons.First, research in the field of applications for remote sensing and estimation of forest ecosystems and wildfires has prominently increased, especially in recent decades due to a rapidly growing number of Earth observation data sources.The progress of forest BA, BS, and post-fire recovery estimations from coarse resolution satellites to high spatial and temporal sensors was studied in the review.Second, during the last two decades, there were several comprehensive reviews from different parts of the world demonstrating a growing interest in the field of study among the professional community [39][40][41][42][43][44].
In order to perform the systematic review, we adopted the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) framework [45].The procedure of identifying, selecting, and analysing relevant articles based on PRISMA is given in Figure 1.The presented schematic diagram follows four phases: identification, screening, eligibility, and inclusion.The PRISMA 2020 checklist (https://www.mdpi.com/editorial_process#standards,accessed on 19 September 2022) comprises a report of several sections with a title, abstract, introduction, methods, results, discussion, and other information.Each PRISMA section consists of specific formulated questions to be identified and analysed in the selected research articles.
Following that, a representative set of keywords were defined for the initial search in the electronic databases Scopus (Elsevier) and Web of Science (Clarivate Analytics): "remote sensing", "forests", "post-fire", "burnt area", "burn severity".A further search was carried out in the same databases using alterations in the keywords for specific satellites (MODIS, Landsat, Sentinel, etc.).The initial attempt showed that 1140 scientific publications were identified consistently with the designated search criteria, of which 660 papers were identified through the Web of Science, and 480 were found in the Scopus database (Figure 1).During the screening phase, the articles search was refined by removing the duplicates from the two databases.Further selection was made by reading the abstracts of the papers to identify research focusing on remote sensing techniques for monitoring and estimation of post-fire forest areas.If the abstract did not clearly describe that the study was related to forest BA, we continued to review the full text to identify the objective and methodology of the article.We also excluded non-peer reviewed publications, and articles where fires have affected other ecosystems (grasslands, shrublands, agricultural lands, meadows etc.).However, this was not always possible due to the fact that, in some regional and global studies, forests were classified along with many other land cover types.

Data Compilation
Finally, 329 papers were shortlisted according to their suitability to the review and fitting of the search criteria, and organized in the reference management Mendeley platform (https://www.mendeley.com/,accessed on 10 May 2022) to conduct further metaanalysis.Details from the selected pool of publications were then extracted and compiled in several different categories related to geographic region (FAO Global Ecological Zone) of the study area, sensor used, spatial extent of the forest fire, temporal scale of the study, type of burnt forest (main tree species), classification and modelling approaches, vegetation indexes, accuracy method used, year of the publication, and reference (validation) data.The complete list of described categories for each selected paper is provided in the supplementary material (Table S1).

Publication Details
The number of studies conducted in the field of remote sensing of forest BA, BS, and post-fire recovery has increased substantially, as is evident from the number of articles published between 2000 and 2020 (Figure 2).This is well-supported by the general trend of increased publication activity in the field of remote sensing, availability, and open access policy for the Earth Observations data (https://www.earthobservations.org/open_eo_data.php,accessed on 10 May 2022), and the importance of the prevention of wildfires for sustainable forest management actions [1,3].Within the decade 2000-2010, we selected 122 (38%) publications for further review, while during 2011-2020, the number was almost doubled, with 204 (62%) publications.

Spatial Extent and Geographical Distribution
The reviewed articles were categorized into five classes of spatial extent: local, multilocal, regional, national, and global scale.In this review, we define local scale as the level of forest unit (national park, forest district, forest reserve, etc.).Locations distributed across a country, different countries or continents without a specific spatial extent are referred to as the multi-local level.The regional level includes research at multiple locations within or across several provinces (states) of the country or cross-border studies in several countries.The national spatial level examines burnt forest within the state boundaries of the study country.Studies at the worldwide level were categorized as global scale.The statistical analysis indicated that the majority of the reviewed articles were conducted on the local (48%) and regional spatial scales (28%) (Figure 4).We identified 30 national (9%) and 28 multi-local (9%) scale studies of forest burnt area with the use of remote sensing, as well as 18 global studies (6%).

Sensor Type
In the study, we also analysed how often particular sensor types and satellite systems were used in the estimations.The reviewed studies were based on various applications of Earth observation data received from 30 different remote sensing sensor systems.The dominant applications were using optical sensors (264 cases, 80%), followed by thermal (Modis) (21%), LiDAR (7%), synthetic aperture radar (SAR) (5%), and aerial photo data (2%) (Figure 6a). Figure 6b shows the distribution of cases that use single or multiple remote sensing data sources.Almost half of the reviewed studies (154, 47%) used a single data source in their studies.In order to enhance the accuracy of forest burned area, another 41% of the studies used multiple optical data sources.About 12% of the studies used a combination of optical data and data from active sensors such as ALOS PALSAR and LiDAR.
In early 2000, the use of AVHRR, MODIS, IRS, SPOT/HRVIR, and Landsat to detect hotspots based on dNDVI images has been widely applied to estimate BA in many studies [69][70][71][72][73][74][75][76][77][78][79][80][81][82][83].Confirmed hotspots are then used to derive regional-level NDVI difference thresholds for BA mapping.As an alternative to the NDVI, a non-linear global environmental monitoring index (GEMI) [84] provided better results for discriminating burned areas in several forest ecosystems [85][86][87].The GEMI was specifically designed to minimize problems of contamination of the vegetation signal by extraneous factors to monitor recently burned areas.Another non-linear burnt area index (BAI), which integrates the minimum reflectance of burnt vegetation in the NIR and the maximum in the red bands, was initially developed for BA discrimination in NOAA-AVHRR images [88].The BAI was adapted to MODIS NIR and SWIR bands (BAIM) for mapping recently burned areas in Mediterranean ecosystems [89][90][91].The mid-infrared bi-spectral index (MIRBI) [92], which applies to SWIR wavelengths, also showed strong spectral separation between burned and unburned areas [93,94].Some researchers found that the MIRBI proved to be reliable to reduce confusions with mixed pixels found in water-land borders and areas affected by previous wildfires [28,95].
Several global BA detecting products have been developed over the past decades based on medium-and high-resolution sensors.Specifically, satellite data from the VEG-ETATION (VGT) sensor aboard the European SPOT-4 satellite were used to assess global BA based on large-scale vegetation features or seasonal burning activity in different ecosystems (boreal, tropical, wooded grassland, savanna, etc.) [96][97][98].The GLOBSCAR project was initiated in 2001 as a part of the European Space Agency (ESA) Data User Programme, with the objective of producing global BA maps on the base of the Along Track Scanning Radiometer (ATSR-2) instrument onboard the ESA ERS-2 satellite [99].The Globcarbon BA product (ESA project), which was developed at a spatial resolution of 1 km, 10 km, 0.250 m, and 0.50 m with monthly temporal resolution to focus on ten complete years (1998)(1999)(2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007), can be flexibly applied to existing archives and readily merged into Earth system models [100].
Two NASA BA products, MODIS MCD45A1 [101] and MCD64A1 [102], which map the spatial extent and approximate date of biomass burning worldwide at a spatial resolution of 500 m, have been validated with the use of Landsat scenes in southern Africa, western United States, and central Siberia.Both products were used to conduct a broad range of studies concerning forest fire estimates and biomass burning emissions [103][104][105][106][107][108][109][110][111][112].The fire information for a resource management system (FIRMS) is also a useful product developed by NASA, delivering MODIS active fire data to forestry managers in more than 90 countries [113,114].A new FireCCI50 BA product was developed within the European Space Agency's (ESA) Climate Change Initiative (CCI) programme [115].The product was generated from the MODIS (2001-2016 archive) red (R) and near-infrared (NIR) reflectance and thermal anomaly data, providing the highest spatial resolution (250 m) among the existing global BA datasets.
In recent years, approaches to planetary-scale geospatial analysis of BA have emerged by using a cloud-based Google Earth Engine (GEE) platform with several satellite data catalogues at different spatial resolutions (MODIS, Landsat, and Sentinel-2).In order to provide higher accuracy and detect small, burned patches, a novel 30 m spatial resolution Global Annual Burned Area Mapping (GABAM 2015) technique was proposed based on a huge catalogue of Landsat time series imagery, as well as the high-performance computing capacity of GEE [116,117].Cross-comparison with the Fire_CCI50 BA product showed a similar spatial distribution and a strong coefficient of determination (R 2 = 0.74) between the burned areas of the two products, particularly in coniferous forests [117].Another version of a global automatic BA algorithm at medium spatial resolution of Landsat or Sentinel-2 reflectance images was also developed in GEE [118].The proposed automatic algorithm has shown globally spatially and temporally coherent results and higher accuracy in the detection of smaller BA.
The growing number of global satellite-based BA products available for the international remote sensing community has led to the development of methods and validation protocols to estimate their accuracy metrics.Although over the last two decades there has been extensive research on the topic [47,[119][120][121], the basis of global BA accuracy assessment has been built according to the "good practice" guidelines for land cover validation procedure [122].The validation protocols for these products have not been formally produced.Therefore, the procedure and methodology rely on previous research publications and validation exercises [110,123,124].Several studies compared the accuracy of the above global BA products, proving that there are substantial differences between them in many regions of the world [47,[125][126][127][128][129][130][131].Six global burned area products were compared using the same protocols and reference data [132].While the overall accuracy (OA) in this study exceeded 99% for all the products, the BA commission error ratio was above 40% for all products and the omission error ratio was above 65% for all products.A Bayesian algorithm (BY-MODIS) to detect BA, originally developed for the North American boreal forests, was compared with MODIS database BA products MCD64A1 Version 6 and MCD45A1 Version 5.1 for the 2002-2012 period in forest ecoregions of Spain [87].The total BA distinguished by the BY-MODIS classifier was 78.6% according to the data obtained from the Landsat images, with the lowest average commission error (11%) out of the three products and a correlation (R 2 ) of 0.82 [87].Meanwhile, the commission and omission errors of the GABAM-2015 machine learning classification were 13.17% and 30.13%, respectively, and the overall accuracy reached 93.92% according to preliminary global validation [116].The GABAM product also detected the largest forest BA in Amazonia, followed by TREES (a regional BA product, developed particularly for the Amazon region), MCD64A1, and Fire_cci [133].The above studies show that there is a need for global and regional products with higher spatial resolution to provide the most accurate BA estimation.
Among the analysed publications, there has been considerable interest regarding satellite-based assessments of BS pattern heterogeneity within the border of a forest fire.For the purposes of the present study, we defined BS as the degree to which an ecosystem has changed due to both short-and long-term wildfire effects [134,135].Relative magnitudes of BS are often expressed in terms of specific forest ecosystem characteristics (vegetation, litter, and soil) that are visually ranked after fire occurrence [136,137].In the context of remote sensing, BS usually refers to measurements made directly after the wildfire, including metrics such as canopy consumption and tree mortality [134] or consumption and charring of organic soil profiles [137,138].
The assessment of burn severity in the field is often based on direct measures of the composite burn index (CBI), which was designed to estimate severity in situ by rating the average burn condition on a field plot [134,139].During the last two decades, CBI has been frequently used in many studies (Table 2) for validating different remotely sensed products and for comprehensive assessment of BS in a wide range of various forest ecosystems [32,[140][141][142][143][144][145][146][147][148][149][150][151][152][153][154].A drawback of the CBI is that it requires knowledge of the pre-fire conditions of the assessed area [136,141], and it does not account for differences in the fractional cover of each ecosystem stratum [155].De Santis and Chuvieco [156] proposed a modified geometrically structured composite burn index (GeoCBI) to improve the retrieval of burn severity from remotely sensed data.It is an operational tool to visually assess the burn severity by fire on the ground in five forest strata (soils, understory vegetation, mid-canopy, overstory, and dominant trees).Several studies [157][158][159][160] reported that GeoCBI is more strongly correlated with spectral reflectance of vegetation on burned areas.Other studies have suggested that burned area and burn severity patterns can also be quantified using remotely sensed change in LAI and empirical models [161][162][163][164][165][166].
While NDVI is still used in current research, during 2005-2020 (see Table 2), the Landsat normalized burn ratio (NBR) has emerged to be meant for BS assessment of forestry wildfires using remote sensing technology [155][156][157][158][159][160].The NBR (and difference NBR) incorporates information about the spectral changes at the surface using the differing responses of near-infrared (NIR, 0.76-0.90mm) and mid-infrared (SWIR, 2.08-2.35mm) bands to generate a scaled index of BS [134,[167][168][169][170][171].Calculated by subtracting the postfire NBR from the pre-fire NBR, the dNBR approach produces a more accurate assessment of burn severity than the single NBR because it is based on the measurement of per pixel changes in reflectance, making the overall approach more transferable across ecosystems [172][173][174][175][176][177][178][179][180].Delineating the calculated dNBR Landsat images and thresholding them into severity classes has formed the basis for the Monitoring Trends in Burn Severity (MTBS) program, whereby all wildfires larger than 200 ha are mapped at the national level [181][182][183].Both the dNBR and dNDVI indices were accurate in detecting high severity in north Patagonian forests, but to a lesser degree for detecting moderate and low severity burns [184].Epting et al. [185] evaluated 13 remotely sensed indices (single bands, band ratios, vegetation indices, multivariate components, normalized differences) for mapping burn severity in interior Alaska, and concluded that post-fire NBR gave the best correlations with the field-based CBI in closed needleleaf, mixed, and broadleaf forest classes, followed by the dNBR.However, they warned that the relationship between the NBR and CBI was poorer in areas with sparse tree cover (open woods, scrub, herbs, and pastures).
dNBR provides a measure of absolute change, whereas RdNBR was developed as a relative measure to account for spatial variation in the pre-fire leaf area index and to improve estimations in areas with heterogeneous and non-productive vegetation cover [186][187][188].There is no common consent among researchers regarding the use of each BS index in terms of forest ecosystems, terrain, and estimated scales.Some studies have concluded that RdNBR provides a more consistent measure of burn severity with higher classification accuracies than dNBR [189,190].Other researchers have reported that the relationships between the NBR and its derivatives (dNBR and RdNBR) to burn severity are highly variable and inconsistent across the burned areas and ecosystem types (savanna, boreal and tropical forests) [191][192][193][194].The study in southeast China showed that the dNBR is a more reliable index for wildfire burn severity mapping than the RdNBR, which resulted in higher R 2 and smaller RMSE values for the forest area [195].Based on field data from six fires, a study in the Canadian National Parks revealed that dNBR had higher overall classification accuracy than RdNBR, and that fires with the lowest forest canopy closure and heterogeneity ranking showed no improvement with RdNBR [167].The results of this research show that the RdNBR-derived model did not estimate burn severity more accurately than dNBR (65.2% versus 70.2% classification accuracy).In addition, the use of NBR or dNBR to assess burn severity may be difficult because of variations in topography and solar elevation angle over the course of the growing season [31,194].The relativized burn ratio (RBR), which provides an alternative to dNBR and RdNBR, was proposed as a Landsat-based metric and evaluated for the wildfires in the conterminous western USA [196].Their results show that the RBR outperformed the dNBR and RdNBR in overall accuracy and in correspondence with field-measured CBI.However, it has also been found out that in heterogeneous areas, with a high presence of shrubs and grass, there were minor differences among dNBR, RdNBR or RBR in estimating BS [197][198][199].While all these standard spectral indices remain the most appropriate for estimating burn severity, there is no strong agreement among researchers regarding the strengths and weaknesses of each NBR-derived index [200][201][202].
Several studies have proposed the use of changes in land surface temperature (LST) and land surface albedo (LSA), as land surface biophysical parameters describing the balance of water and energy, for estimating the BA and BS of forest ecosystems [144,[203][204][205][206][207][208][209].All of the studies reported the expected temperature increase in the immediate post-fire environment.For example, NBR and three indices in which the NBR is enhanced with LST revealed the best performance in assessing burn severity in chaparral shrublands [210].Maffei et al. [211], based on the time series of daily Aqua-MODIS LST data (2003-2017) and the harmonic analysis of time series (HANTS) algorithm, demonstrated that the LST anomaly is a covariate of both burned area and fire duration.The comparison of the above publications in relation to the burn severity may aid in better understanding of forest burn patterns at different spatio-temporal scales, as well as providing reliable information for forest managers' systematic monitoring of wildfires and restoration efforts.

Active RS Sensors
During the last decades, several techniques have been developed for using synthetic aperture radar (SAR) and LiDAR data for BA mapping and fire damage assessment in forested areas [125,149,163,212,213].Unlike optical sensors, the radar data can be obtained around the clock, at any position of the sun and adverse weather conditions, including rain, smoke, and clouds, which is crucial in emergency situations.SAR uses microwave energy for BA mapping, given that the backscatter signal is sensitive to vegetation structure and forest biomass [214,215].Radar-based measurements, such as polarimetric decomposition coefficients, interferometric coherence, and backscatter intensities, show satisfactory BA estimation accuracies in wildfire-affected forests [134].The detection of BA from SAR is based on ecological changes that occur after wildfire, including the removal of tree canopies, exposure of rough ground, and increased water content in the burned surface [216,217].These changes are directly proportional to the degree of BA, appearing on SAR images as relatively darker objects compared with non-affected forest areas, particularly when rainfall occurs after the fire [218].
With the launch of ESAʹs Sentinel (S) satellite, SAR images are now easily accessible with systematic acquisitions at a 5 × 20 m spatial resolution, which largely increases the capacity of BA detection [219,220].The combined use of the S1 SAR and S2 Multi Spectral Instrument (MSI) sensors made it possible to detect 36,000 ha of BA in Marantaceae forests of the Congo Basin, characterized by open tree canopy and a tall herbaceous layer [27].A new approach in the use of Sentinel-1A and RADARSAT-2 data from different orbits demonstrated their good agreement and potential of a thresholded, multi-temporal, αangle-based index (NDαI) for mapping BA in a fire-prone region of the Western Cape (SA) [221].The study also revealed that large omission errors, correlating strongly with areas of high local incidence angle, could be minimised by the combined use of both datasets.Another study of BA detection and mapping in tropical and sub-tropical Africa based on two Sentinel-1 and one Sentinel-2 algorithms suggested that optical-based algorithms provide a significant increase in accuracy compared to radar-based algorithms [131].
Light detection and ranging (LiDAR) has recently emerged as a powerful tool for producing 3D forest structural attributes, mapping BA, and assessing post-fire regeneration [222][223][224][225].Some studies have combined multispectral/hyperspectral data and LiDAR data to estimate BS patterns using analysis of tree height and forest canopy structure [226][227][228][229][230][231].There has been only a few studies estimating the spatial-temporal match of pre-and post-fire LiDAR acquisitions to quantify BS on forest structure.Bishop et al. [232] assessed the accuracy of classifying colour-infrared imagery in combination with post-fire LiDAR, and with differenced (pre-and post-fire) LiDAR, in a second-growth coast redwood forest, CA.The research demonstrated a moderate increase in overall accuracy compared to the use of imagery alone; the model using NDVI and post-fire LiDAR was 85% accurate.Montealegre et al. [233] developed a methodological approach combining post-fire airborne LiDAR data and field assessed CBI to estimate BA in Spanish forests.Developed regression models confirm that classified maps of BA are acceptable up to 85.5% overall accuracy.Liu et al. [234] investigated the application of photon counting LiDAR data from NASA Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) to classify BA over forests in northern California and western New Mexico.The classification results indicate 83% overall accuracy of BA maps based on spaceborne LiDAR, Sentinel-2, and Landsat 8 derived samples [234].
While using airborne LiDAR data provides good potential for assessing BA and BS in forestry with high accuracy, nowadays, the application of these data over large areas is limited by its cost associated with data acquisition.Another drawback to the incorporation of LiDAR measurements for the purpose of BA estimations is the lack of up-to-date pre-fire data for most forestry areas.Nevertheless, the use of LiDAR in wildfire assessments is expected to increase based on the capability to provide high spatial resolution data.

Monitoring of Post-Fire Forest Recovery
Quantifying post-fire effects become increasingly significant for understanding how forest ecosystems respond, including vegetation recovery, patterns of succession, ecosystem resilience, ecological impact, and planning forest management [40,235,236].In this respect, RS has been recognized as an indispensable tool for monitoring reforestation patterns after wildfire [237][238][239][240][241].Among different quantitative techniques, Tasseled Cap [242] and Multitemporal Kauth Thomas transformations [243] are of particular interest to the researchers as they are based on assessing satellite imagery of three key indicators of post-fire vegetation conditions-brightness, greenness, and wetness.Numerous studies have used a comprehensive analysis related to both transformations for mapping changes and disturbances in forests after wildfires [244][245][246][247][248][249][250].
Several studies on post-fire dynamics have been carried out in boreal forest ecosystems using optical sensors.Hicke et al. [251] estimated a mean NPP post-fire recovery for North American boreal forests of about 9 years using a 17-year time series of AVHRR NDVI observations.Another study used two NDVI AVHRR 17-year time series datasets to investigate vegetation recovery after wildfire in the boreal forest of Canada [252].The results indicated that the rate of forest recovery varied in three episodic wildfire years (1981, 1989, and 1995), and that recovery times to preborn levels (5 years or more) were shorter than in previous studies.Cuevas-Gonzales et al. [253] observed spatio-temporal patterns of the fraction of absorbed photosynthetically active radiation (fAPAR) with the use of the MODIS time series to estimate post-fire dynamics during and two years after a fire event in Siberian boreal forests.Their results indicated that the differences in the fAPAR trajectories are correlated with the forest type, and fAPAR changes are not significantly different among the boreal latitudinal regions.The post-fire regeneration of Siberian boreal forests was studied using a 16-year time series of NDVI and NDSWIR (normalized difference shortwave infrared index) derived from MODIS [35].The results have proved that more than 13 years are needed for the temporal NDVI and NDSWIR signals to recover fully after wildfire.Epting and Verbyla [254] used Landsat TM/ETM+ images to estimate forest succession over a 16-year post-fire period in interior Alaska.The study results indicated that patterns of post-fire Alaskan boreal forests are strongly linked to both the types of pre-fire vegetation and BS.In sixteen years after the fire, most of the high BS areas with closed needle-leaf, open needle-leaf, and mixed forest classes shifted to woodland or shrubland [254].Frazier et al. [255] used three spectral recovery metrics derived from an annual Landsat based per-pixel NBR time series to estimate the recovery cycles of Canadian boreal forests.They found that post-fire spectral forest recovery rates have increased during 1986-2011 by 18% in the Taiga Shield East and 9% in the Taiga Shield West of Canada.
Many studies of post-fire vegetation responses are based on the discrimination of spectral bands and vegetation indices (mostly NDVI, dNBR, and EVI) by MODIS, Landsat, SPOT, and Sentinel multitemporal imagery in different regions and forest ecosystems of the world [256][257][258][259][260][261][262][263][264][265][266][267].Numerous studies have observed that NDVI has a higher correlation with post-fire vegetation regrowth, leaf area, and biomass than other VI [36,135,268].Severe changes in the ecosystem growth conditions caused by forest fire often result in rapid reforestation processes that can be observed at the regional level with MODIS 250m NDVI time series data [35,36,[269][270][271][272].Ryu et al. [273] used MODIS and Landsat TM/OLI data to evaluate forest recovery processes and physiological activity after the April 2000 forest fire on the eastern coast of South Korea.The study results show that the pine (Pinus densiflora) ecosystem conditions improved to the initial state 8 years after the fire event, while vegetation biomass required more time to recover.Meng et al. [274] also used changes in NDVI following thirty-five fires in mixed conifer forest in the Sierra Nevada Mountains of California between 1999 and 2006.They reported that post-fire drought has a significant effect on the recovery of burned mixed-conifer forest, while burned red fir forest was sensitive to post-fire temperature.Lee and Chow [275] found that vegetation rapidly recovered in the first three years (NDVI increased substantially) following the wildfire in Bastrop State Park (USA), but that recovery was still significantly delayed compared to the control plot of the field data.On the other hand, the NDVI curves performed differently when estimated by BS level, showing a net decline in heavy and moderate BS areas [275].
Post-fire recovery dynamics were assessed in Mediterranean pine (Pinus halepensis and Pinus pinaster) forests (Spain) using temporal segmentation of the Landsat time series  and trajectory metrics from Tasseled Cap Wetness (TCW) and Tasseled Cap Angle (TCA) [276].The research suggests that pine forest recovery rates were highly correlated to post-fire climate in the mid-and long-term, and with BS in the short-term [276].Several studies in modelling of forest recovery and regeneration rates adopted other significant factors such as vegetation type, underlying topography, soils, slope-aspect, rainfall, and drought indicators [277][278][279].Zhao et al. [280] estimated post-fire forest spectral recovery in the Greater Yellowstone Ecosystem (GYE) using the Landsat Time Series Stack-Vegetation Change Tracker (LTSS-VCT) algorithm and high spatial resolution images from Google Earth.They found that forest regeneration is highly correlated with forest type, elevation, and soil type of the GYE.Another study in Mediterranean forest ecosystems [281] demonstrated evidence that the number of ongoing wildfires within a short period decreased the forest recovery negatively measured as EVI, which was also influenced by micro-climatic effects (soil and environmental humidity).Such studies also suggest that results are more meaningful when other important factors are considered, modelled, and compared with the forest recovery on BA.
In addition to optical sensors, the value of using unmanned aerial vehicles (UAV), SAR, and LiDAR imagery for estimation and monitoring of post-fire forest recovery has also been demonstrated [282][283][284][285]. Padua et al. [286] reported that the efficiency of Sentinel-2 satellite data for post-fire forest regeneration monitoring in the north-eastern region of Portugal is more cost-effective than UAV data.Several studies confirmed that the combination of SAR and Landsat bands was essential for estimation of regeneration stages in different forest ecosystems [282,287,288].Polychronaki et al. [259] estimated post-fire vegetation recovery on the Mediterranean island of Thasos using multi-temporal SPOT and European Remote-Sensing (ERS) (C-band VV) images covering the period from 1993 to 2007.Results showed that the use of the ERS backscatter coefficient reduced the commission errors almost by half when mapping the forested areas recovery in Thasos [259].
LiDAR stands out as an important technology capable of being directly associated with ecosystem biomass and post-fire changes on the local level, since it measures the three-dimensional structures of forest cover.Sato et al. [289] estimated post-fire changes in forest canopy height and biomass using airborne LiDAR in western Amazonia.Their results indicated that even ten years after the occurrence of an understory fire event, burned forests had significantly lower biomass and height than control sites.Gordon et al. [290] investigated the effect of mixed severity wildfire on the regrowth of open forests in Australia on the base of LiDAR data and field surveys.This study showed that vegetation recovery of Eucalyptus, Callitris, and Acacia species was very fragmented.The regrowth of mid-story vegetation was 22-40% lower in BA of low and moderate compared to high fire severity [290].In order to quantify how wildfire impacts forest recovery, several researchers have examined the combined use of optical sensors imagery and LiDAR data [225,226,230,291].Meng et al. [292] studied the recovery rate of mixed pine-oak (Pinus rigida) forest in Long Island (NY) three years after wildfire by leveraging 1 m of simultaneous airborne imaging spectroscopy and LiDAR and 2 m of satellite multi-spectral imagery.They presented a new method for the estimation of tree species (oak vs. pine) recovery rates to different BS levels, derived from large spatial-temporal scales.According to the study, pine trees were more resilient to high BS with a higher recovery rate (12% per year) than oak (Quercus alba and Quercus coccinea) species (4% per year) [292].Bolton et al. [293] estimated vegetation recovery in the Canadian boreal spans following wildfire, combining LiDAR, MODIS GPP, and a national Landsat-derived record of forest change.The research showed difference in forest canopy structure across the boreal spans, while correlations to GPP were stronger for forest stands by 15-25 years post-wildfire.
Combining different RS images (optical, SAR, and LiDAR) with BS maps provides a more comprehensive assessment of post-fire forest recovery patterns.Therefore, further studies based on multi-source analysis are still needed with the inclusion of structural forest ground variables, species composition, environmental conditions, vegetation indexes, and BS [42,44,46].
MESMA/SMA are the most commonly used techniques for burned forests and BS detection and classification.SMA is based on modelling image spectra of the combination of particular endmembers (such as soil, forest, non-forest, etc.), and can reduce mixed pixel effects arising in the visual classification and digital processing of RS data [71,137,310].SMA resolves the concern associated with BA that it is smaller than the pixel size of an estimated image [40,308,309].Compared to the typical linear SMA technique, MESMA enables the use of multiple endmembers within-class spectral variability and overcomes the limitation of testing the same number of endmembers for each pixel in the image [37,239].Many researchers used SMA/MESMA for mapping burn severity and post-fire vegetation monitoring [154,228,231,239,274,[311][312][313].

Classification Methods
Since 2000, many advanced image classification methods have been developed to extract and detect the BA information from remotely sensed imagery.We summarized the classification methods into four main groups -Supervised (pixel-based and geo/objectbased image analysis (GEOBIA/OBIA)), Unsupervised (pixel-based), Hybrid, Manual, and Other.The majority of remote sensing BA studies have shown preference for the supervised classification approach (82%) (Figure 8a).This is mainly because the supervised method ultimately deals with many classification algorithms, while the unsupervised method employs only a few of them, which usually have lower accuracy than the supervised methods.Several studies used supervised parametric classification algorithms such as maximum likelihood (ML) (11.3%) [53,173,210,297,[314][315][316][317][318] and Mahalanobis distance (MD) (3.8%) [319] (Figure 8b).
In the last decade, the GEOBIA method has emerged based on image-objects segmentation and classification rather than pixels, which increased the land-cover studies with the use of high spatial resolution remote sensing imagery [369].Of the total BA studies, 9% used OBIA/GEOBIA classification with the application of ground truth data and regional forest maps to validate the classification accuracy [370][371][372].Application of the GEOBIA improves detection and fuzzy logic classification accuracy of BA [147,215,259,278,373,374]. In addition, the object-based approach, despite the inherent computational costs, was reported to be less time-consuming than the pixel-based approach [299].However, pixel-based approaches are still more widely used by researchers.

Classification Accuracy
Several metrics such as the Kappa coefficient, overall accuracy, and omission and commission errors are typically used with the confusion matrix for burned area classification performance evaluation.A total of 151 studied papers used the error matrix method, regression analysis, and modelling for validation of BA estimation results.Table 3 summarises the distribution of the reference data type and validation methods used in these studies.Out of 151 papers, 49 (32%) used ground truth data (field survey, composite burn index, GPS points, and forest inventory sample plots) for reference purposes.From these, 20 papers provided Kappa coefficients and overall accuracy together with the error matrix results.The second group of 31 studies was based on the reference data retrieved from high resolution imagery (e.g., Google Earth, WorldView, aerial photo).Seventeen studies of this group used the coefficient of determination (R 2 ) with/or root mean square error (RMSE) value to describe the possible prediction accuracy of the efficient models.Thirty-seven studies used ancillary data produced by the government authorities (e.g., forest inventory maps, land use/land cover, official data of statistical agencies).In another thirty-four studies, the validation and assessment of forest BA were based on the combination of the above-mentioned reference data.For example, some studies used ground truth data and high-resolution images or ground truth data with local maps.The results in Table 3 indicate that though during the estimated 20 years the metrics of confusion matrixes were still in high demand among researchers, regression analysis and modelling (R 2 /RMSE) were the most used methods in accuracy estimation and mapping of the forest BA.
For a comprehensive analysis of the RS-based forest BA studies, the reported overall accuracies were measured and compared with several factors, such as the year of publication, classification method, respective sensor, and the reference data considered in the validation of classification results.The classification accuracy of estimates for forest BA varies significantly among the assessed studies in all four groups of factors, especially with respect to the classification accuracy of different sensors and reference data (Figure 9).
From the box and whisker plots on Figure 9a, it can be seen that throughout 2000-2020, the median of overall accuracies were gradually improving.In 2000-2005, the median overall accuracies of BA classifications were about 85%, while in 2016-2020, the median increased sufficiently up to 94%.This positive tendency is likely a result of improved classification methods used in the evaluated studies, including novel classification techniques, advanced sensors, and the high quality of reference data.The most accurate results of the forest BA mapping were obtained using RF, SVM, and MLH classifiers (OA about 90%), followed by OBIA, ANN, and an unsupervised group of classifiers (OA about 88%).DT and SMA methods are characterized by a similar median classification accuracy of 85.8% and 86.4%, respectively (Figure 9b). Figure 9c with the box and whisker plots shows the median overall accuracies of nine different sensors.It is observed that, except for RADARSAT images, the median classification accuracy is generally above 85%, with a range from 76.8% to 98.8%.IKONOS demonstrates the highest accuracy, at 95.9% median overall accuracy of BA classification, followed by Landsat TM (92.5%),SPOT (91.2%),MODIS (90.9%), and Landsat ETM+ (90.2%).Figure 9d shows that among the reference data used for validation, the median overall classification accuracy in BA studies is highest for the ground truth data and combined data, at about 90.5%.A similar median classification accuracy of 88% is also observed for ancillary data and high resolution images.The advances in RS-based forest BA mapping methods described in the reviewed papers also leads to a great variety of overall classification accuracies.In the case of approaches using input from several factors (e.g., combination of sensors, ancillary and ground truth data), this might have the potential to improve the accuracy and interpretation of classification results.In turn, the accurate maps of BA might provide a more reliable interpretation of forest structure relative to the modelling of post-fire patterns and recovery processes.

Research Perspectives
The meta-analysis results provide a base for the above-mentioned challenges for future research in the field of forest BA, BS, and post-fire recovery:  It is important to conduct more research in the forest ecosystems of South America, Africa, and Eurasia, because these regions have a great impact on the global balance of carbon and climate change.


Further identified investigations are likely to include the use of higher resolution data, algorithms for multi-criteria analysis, empirical models, and data fusion from multiple sensors.The adoption of these methods and data can potentially address the misclassification and high variability in BA, BS, and post-fire recovery studies.


Neural networks, classification and regression trees (CARTs), fuzzy modelling, and OBIA are also very promising in future studies with the application of big data algorithms and emerging processing platforms.The use of machine-learning algorithms and cloud computing techniques (e.g., GEE) provides a new opportunity for the field of study.


The review suggests that more research needs to be done on finding methods for reducing RS image classification errors (omission and commission) and increasing global and regional BA map accuracy, particularly in data-poor regions.


The combination of RS imagery from passive and active sensors can provide more insight in the field of study, including the possibility to overcome cloud cover issues in a forest BA.


We expect to see an increase in the integration of LiDAR, radar, hyperspectral sensors, and emerging UAS technologies in future BA studies.There is also huge potential for research on the spatial and temporal pattern of post-fire recovery and approaches to forest BA monitoring with the aim of improving ecosystem sustainability.

Conclusions
This paper presents a comprehensive overview on the progress and development of RS-based applications in the detection and monitoring of forest burnt area.We identified a total of 329 relevant papers in a selected 48 peer-reviewed journals pertaining to all aspects of remote sensing of forest BA within the last two decades.A database containing information on RS-based forest BA estimates was constructed to serve as a base for metaanalysis.We analysed the papers with regard to the research focus, ecological zoning, spatial resolution of the RS images, temporal scope, tree species, spectral indices, reference data, overall accuracies, and classification methods applied.A deep meta-analysis was carried out to discuss the application of RS systems in forest post-fire studies worldwide, which differentiates our study from previous reviews.Topics under consideration include burned area estimates and mapping, burn severity assessment, and monitoring of postfire forest recovery.
Our review of the available papers indicates that the number of remote sensing studies conducted on forest BA increased substantially from the early 2000s to reach a peak in the 2011-2020 decade.The dominant studies were using optical sensors (261 cases, 80%) based on Landsat, Sentinel, Terra, Aqua, Sentinel, AVHRR, and SPOT satellites.According to the FAO Global Ecological Zone map, most of the reviewed studies were carried out within subtropical dry forest (21.4%, 92 cases).With regard to tree species discrimination, the studies mainly described pine (37.0%).The NDVI and dNBR were the most frequently used VI compared to other indices.The studies show that there are some challenges, mainly due to spectral mixing and poor validation data quality, in determining BA and BS.The use of multisensor images may improve the accuracy and reduce uncertainty in BA estimations.For instance, combining optical time series at moderate resolutions with radar data provides a good opportunity to improve the overall accuracy of the assessments.There is also a need for global and regional products with higher spatial resolution to provide the most accurate BA estimation.The use of RF, SVM, and MLH classifiers has allowed the detection of forest BA classes with no significant difference among them.During 2000-2020, classification techniques gradually tended to reach higher overall accuracies, specifically when using ground truth and combining data from different reference sources.Consequently, remote sensing is expected to improve methodologies to detect and monitor BA in a fast and cost-effective manner, and thus enhance its role in sustainable forest management.

Figure 1 .
Figure 1.The schematic diagram of the papers selection order from 2000 to 2020 in the field of remote sensing of forest BA, BS, and post-fire recovery in the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) framework.At the eligibility phase, the full text of the remaining 605 articles were carefully reviewed based on the following criteria:  Remote sensing techniques are the main components of the study;  Study should attempt temporally explicit estimation of the forest BA and BS using different datasets for the validation of mapping results;  Paper appropriately describes the methods and/or results of the study so that they can be interpretable and replicated;  Focus is made on the forest stands rather than other ecosystem attributes (soils, water, undergrowth);  Study should report detailed accuracy assessment results.

Figure 2 .
Figure 2. Temporal distribution of available and selected publications from 2000 to 2020 for the review in the field of remote sensing of forest BA, BS, and post-fire recovery, obtained from Scopus and Web of Science databases.Data analysis revealed that 48 peer-reviewed journals published the selected papers.The main source journals of the 329 papers were: Remote Sensing of Environment (74 papers), Remote Sensing (56 papers), International Journal of Remote Sensing (39), International Journal of Wildland Fire (26), International Journal of Applied Earth Observation and Geoinformation (15), and Forest Ecology and Management (14) ranked as the top seven journals that published no less than 10 research articles during the past two decades.The top 20 journals are presented in Figure 3, which together contained 285 publications or approximately 88% of the selected papers.The other 28 journals (12%) that published only two relevant articles were eliminated from the descriptive statistics in Figure 3.Most of them, which do not directly deal with remote sensing, were distributed across a wide range of journals (e.g., Journal of Biogeography, Fire Ecology, Ecological Indicators, Climatic Change, Landscape and Urban Planning, Biogeosciences, Floresta, Ecological Modelling, Science of Total Environment, Global and Planetary Change, Contemporary Problems of Ecology, Izvestiya, Atmospheric and Oceanic Physics, Biological Conservation, Catena, Journal of Tropical Ecology, Environmental Monitoring and Assessment, Geophysical Research Letters, etc.).

Figure 3 .
Figure 3. Top 20 out of 48 peer reviewed journals where the selected papers were published.

Figure 4 .
Figure 4. Distribution of selected articles at different spatial scales.

Figure 5 .
Figure 5. Geographic distribution of study site locations conducted per country: (a) global map of reviewed studies; (b) total number of studies per country (>1).

Figure 6 .
Figure 6.Remote sensing sensor types and sensor systems used in selected publications: (a) number of sensor types applied; (b) sensors combination; (c) distribution of optical sensor system used in the studies.

Figure 7 .
Figure 7. Distribution of studies: (a) according to the Global Ecological Zoning framework; (b) by the main tree species on the burnt sites.

Figure 9 .
Figure 9. Overall accuracies reported in RS-based BA classification studies based on: (a) the year of publication, (b) the classification method, (c) the satellite sensor, and (d) the reference data.

Table 1 .
Tree species on the burnt sites described in more than four cases (156 articles in total).

Table 2 .
Spectral indices used for the review.

Table 3 .
Distribution of the reference data type and validation methods with the respective number of studies.