Different approaches make comparing studies of burn severity challenging: a review of methods used to link remotely sensed data with the Composite Burn Index

Keywords: burn severity, CBI, Composite Burn Index, dNBR, dNDVI, fire severity, GeoCBI, landsat imagery, NBR, NDVI, RdNBR, remote sensing, spectral index.

Estimates of burn severity, here defined as the ecological effects of fire, or fire-caused change (Lentile et al. 2006), inform immediate post-fire recovery projects, future fuel treatments, and long-term ecosystem analysis (Miller and Thode 2007). In the immediate post-fire environment, common priorities are to preserve the soil and maintain the vegetation community (Fernández-García et al. 2018a). A second key function of severity assessments is to understand how fire effects vary with ecosystem composition and structure. Linking pre-fire conditions with potential post-fire outcomes helps managers to design better fuel treatments, such as thinning and prescribed fires (Churchill et al. 2013, 2022; Larson et al. 2022). Severity estimates can also help to predict long-term consequences to a post-fire ecosystem (Macdonald 2007), such as changes to forest succession (Johnstone and Chapin 2006). Managers need predictive models of wildland fire effects on vegetation communities, wildlife populations, and hydrologic function (Sorbel and Allen 2005), and ongoing records of fire effects allow ecologists to assess effects of climate change over time.

Commonly measured fire effects on the ground include consumption of vegetation, destruction of leaf chlorophyll, exposed soil, charred stems, and altered moisture (Epting et al. 2005). These ecological impacts lead to spectral, thermal, and structural changes on the land surface, which can be captured by remote sensors (Epting et al. 2005; Mallinis et al. 2018). Combinations of these post-fire soil and vegetation conditions, including those with multiple ecosystem attributes, are used to assess burn severity (Table 1). Methods may be quantitative or qualitative; all incorporate a choice of fire-related variables that depend on management and ecological objectives as well as sampling scale (Key and Benson 2006). Post-fire site characteristics are typically considered relative to the pre-disturbance environment (Miller and Thode 2007), so severity is also a subjective measurement that can change with the time of observation and information available (Lentile et al. 2006).

Table 1.  Ecosystem attributes used to evaluate burn severity.

The size and variability of wildland fires pose several challenges for characterising fire effects on the ground (Chen et al. 2011). First, accessing remote locations where fires have occurred requires resource-intensive and logistically challenging fieldwork (Morgan et al. 2014). Second, the heterogeneity in burn severity across burned landscape requires many plots to represent the range of conditions in burned areas (Morgan et al. 2014). Finally, many burned areas are poorly represented spatially from field plots alone, due to the sheer size of burned areas relative to the size and number of field plots that are logistically possible (De Santis and Chuvieco 2007).

Remote sensing is a powerful tool to understand spatial patterns across burned landscapes, providing a synoptic view from space- or airborne sensors, (Lentile et al. 2006; Wulder et al. 2009; Soverel et al. 2010) and allowing observation of post-fire effects inaccessible from the ground (Key and Benson 2006; Lentile et al. 2006; Murphy et al. 2008). That said, although remotely sensed data provide extensive spatial and temporal coverage that field observations lack (Schepers et al. 2014), they must be linked to ‘ground-truth’ field data (Morgan et al. 2014). The choice of field data depends on the goals of the study; there is no one-size-fits-all approach to evaluating severity (Keeley 2009). The range of potential research questions and applications of these assessments requires different approaches determined by a combination of management, ecological goals, and field sampling designs (Ryan and Noste 1985).

The Composite Burn Index (CBI) integrates multiple metrics across vegetation strata and soil, which, used together (as a measure of site burn severity), provide an overall scope of fire damage (Key and Benson 2006) (Fig. 1). Individual components of the index may also be considered separately, depending on post-fire management needs (Key and Benson 2006; Zhu et al. 2006; Keeley 2009). CBI has been adapted in many ways, including general modifications such as the geometrically structured CBI (GeoCBI, De Santis and Chuvieco 2009), and site-specific modifications such as adding or omitting specific attributes (Epting et al. 2005; Allen and Sorbel 2008; Hoy et al. 2008; Schepers et al. 2014; Fernández-García et al. 2018b).

Fig. 1.  Hierarchical structure of the Composite Burn Index (CBI).

CBI has emerged as a popular measure of field-based severity in remote sensing for several reasons. First, the rapid protocol allows quick deployment and assessment over large landscapes (Parks et al. 2019). Second, CBI has generally robust relationships with field measures (Saberi et al. 2022) and satellite spectral fire severity indices (French et al. 2008), providing strong support for its use to estimate fire effects across landscapes (Parks et al. 2019). Third, for general assessments of severity, the CBI may be more complete than other classification systems based on single indicators of burn severity (Sikkink 2015).

In addition to the CBI, several alternative composite severity measures have been published in the fire ecology literature. The GeoCBI accounts for differences in fractional cover of each stratum and changes in leaf area index (LAI) for the intermediate and tall tree strata (De Santis and Chuvieco 2009). Fractional cover and LAI describe the influence of vegetation cover on reflectance of different strata within a given plot, and from a remote sensing point of view, the spectral response of a plot is strongly related to the vegetation coverage per stratum (De Santis and Chuvieco 2009). Other modified versions of the CBI include a Weighted CBI (WCBI) (Cansler and McKenzie 2012) and a Burn Severity Index (BSI) (Loboda et al. 2013). The WCBI generally uses the original CBI protocol for sampling severity across strata, but it weights scores by the fraction of cover for each stratum. Although the method is similar to the GeoCBI in that regard, specific weights are study dependent. The BSI was introduced by Loboda et al. (2013) as a less rigorous approach that mimics CBI plots but saves time in the field.

Many field measures of severity were designed to scale across broad landscapes using remotely sensed data, based on the underlying ecological causes that drive changes in spectral reflectance between pre- and post-fire landscapes (Supplementary Table S1). The ecosystem attributes characterised by the CBI and related composite severity measures span the range of change expected to be captured by remotely sensed data. Overall, studies that use remotely sensed data to link to continuous field measures based on composite severity measures follow a similar methodological framework to each other (Fig. 2). However, a number of key decisions are made at each step. Differences in these decisions result in studies that use many unique avenues, and we have limited understanding on how that may affect findings across studies. The range of choices available for analysis at each step may, in totality, be considered a ‘decision menu’ that depends on the research questions and preferred methodological approaches of the study author(s).

Fig. 2.  Flow diagram showing framework of methods for study design that relate remotely sensed data to continuous measurements of burn severity from the CBI and similar composite severity indices.

To our knowledge, no study has yet summarised the methods and important decision points that link remotely sensed data to CBI as a continuous measure. However, studies such as Stambaugh et al. (2015) demonstrate the potential for methodological decisions to strongly influence the relationships between remotely sensed data and field observations. For example, R² values for the same set of plots ranged from 0.18 to 0.78 depending on whether models used remotely sensed data from initial versus extended assessments, were processed with interpolation or not, and whether they used the delta normalised burn ratio (dNBR) or relativised dNBR (RdNBR). This example highlights the difficulty of comparing model results across studies where methodologies may differ. However, the existing body of literature linking CBI measures to remotely sensed data lacks a comprehensive review of the possible methodological choices that can be made during analysis. This paper sought to conduct a review of the many choices of the ‘decision menu’ and identify the limitations imposed by the lack of consensus in modelling approaches.

The goal of this review was to compare empirical methods used to estimate burn severity measured with CBI both in the field and through remote sensing to understand the impacts of key analytical decisions and research gaps in the existing literature. We did this by exploring:

1. Where and when have studies been conducted?

2. How is burn severity measured in the field?

3. How is burn severity measured from satellites?

4. How are field observations and remotely sensed data modelled together?

Addressing these key questions from our review of 62 studies will improve our understanding of the state of the science regarding sampling of post-fire environments and the ability to map burned landscapes using remotely sensed data. Additionally, the aggregation of research results from the numerous studies that used a variety of analytical methods will provide insights into whether findings may be generalised across studies or are too study specific to interpret in such a manner.

We developed a procedure to guide our process for conducting a systematic literature search, identifying relevant articles, and extracting data for analysis. This review broadly follows the framework developed by Pullin and Stewart (2006) for conducting systematic reviews in conservation and environmental management. Their guidelines for conducting systematic reviews are based on evaluating the effectiveness of management and policy interventions, but we only used their approach to develop a search protocol and to identify and review relevant studies to generate a database from papers that used remotely sensed data and continuous measures of burn severity based on the CBI. Therefore, our work does not investigate the efficacy of a specific treatment (e.g. mechanical thinning or prescribed fire) as the systematic review protocol dictates. Our search criteria were based on fire and field measurements and four remote sensing technologies (optical, thermal, radar, and lidar) (Table 2). For each criterion, a set of keywords was developed for the database queries (see next section).

Table 2.  Search criteria used during review.

The CAB Direct (https://www.cabdirect.org), Environmental Science Collection (https://proquest.libguides.com/environmentalsciencecollection), and Web of Science (https://www.webofknowledge.com) databases were systematically searched for original research that investigated the relationships between remotely sensed data and continuous measures of burn severity assessed with the CBI or GeoCBI. Search terms were combined through Boolean operators so that each query contained the keywords for fire, field measurements, and one remote sensing technology. We used the resulting expression to search abstracts in each database. After combining the results from all three databases, duplicate entries were removed (Supplementary Fig. S1). Any studies that we missed would mainly have been studies published in a journal outside of our database queries (CAB Direct, Environmental Science Collection, Web of Science) or not having the keywords in our search.

Articles were screened to determine whether each entry met the inclusion criteria. Articles were excluded if they (1) did not use CBI/GeoCBI measurements, (2) used simulated CBI/GeoCBI or simulated remotely sensed data, (3) did not analyse CBI/GeoCBI measurements as continuous variables, or (4) were not in English. We focused on the use of CBI as a continuous variable because it provides a more flexible approach than categorical measures that classify measurements into bins such as low, moderate, and high severity. Continuous data can be adapted to multiple different ecological phenomena. For example, different studies may use various threshold breakpoints for different levels of severity depending on the model used and the user’s needs (Boucher et al. 2017), or chosen according to the relevant ecological effects (Hall et al. 2008; Cansler and McKenzie 2012). However, severity values are frequently classified in order to compare multiple fires, analyse spatial patterns, or communicate with managers regarding prioritisation of response (Cansler and McKenzie 2012). Readers may find additional insights into the use of categorical observations of severity in Cansler and McKenzie (2012).

We selected all empirical studies published in peer-reviewed scientific journals through 2019, excluding published literature reviews, dissertations and theses, and government reports. All articles matching the review criteria were retained regardless of the study’s location or ecosystem type. Four additional articles were found through searching the literature cited of the papers included in our search. The articles that met our inclusion criteria were reviewed to extract the relevant information, which was entered in a rubric format. The data for each paper were captured in individual databases for information at the article level (Supplementary Table S2), fire level (Supplementary Table S3), and comparison level (Supplementary Table S4). We compiled 62 studies based on the structured database queries; a list of studies included in this review and the number of field observations to remotely sensed data comparisons extracted for analysis are included in Supplementary Information S2.

To locate fires examined by studies included in this review, we used a combination of georeferenced study maps, the published manuscripts, provided coordinates, or ancillary databases (Sikkink et al. 2013; Picotte et al. 2019; CAL FIRE 2020; MTBS 2021). Fires were mapped for cases where we could confidently identify the study fire based on name, location, size, number of field plots, or other identifying characteristics. The spatial mapping of fires allowed us to overlay terrestrial ecosystems of the world (Olson and Dinerstein 2002) and identify which ecosystems may be more or less understood. For fire size, information was found either in the study itself or through ancillary databases.

Although our focus was on the use of burn severity as a continuous variable, many included studies that used continuous data in their analysis also reported thresholds used for classifying burn severity. In such cases, we recorded the distribution of field plots classified by unburned, low, moderate, or high burn severity. These data were assessed in two ways: (1) by summing the number of plots per severity category across the studies; and (2) calculating the proportion of plots placed into different burn severity classes. By assessing classified data, we were able to present some results on the variability of the threshold values used to bin continuous observations. Additionally, because many studies did not show a distribution of continuous plot severity scores, we used this subset of studies that provided classified distributions to gain some understanding of how plot severity varied across studies.

To assess the timing of field campaigns, we used the date of ignition for the fire and date of field sampling. The smallest temporal grain we could identify for cases was to the nearest month. Results were limited to studies that reported both the timing of fires (or studied fires that where ignition date was available elsewhere) and the timing of field campaigns.

In some cases, we contacted authors to ask for information not provided in published manuscripts such as field plot size and timing, index value extraction methods, etc. All summary statistics and figures – with the exception of the study map – were conducted using the R programming language (R Core Team 2017). Where presented, standard deviations were calculated using the population standard deviation equation (Zar 1999). Due to missing information, some summary statistics and analyses are based only on cases where data were available. The number of studies, fires, or comparisons for each analysis are presented in each relevant table or figure caption or in the figure itself.

The systematic search resulted in 62 papers published in 2004–2019. All used field measures of burn severity based on the CBI protocol (CBI, GeoCBI, WCBI, BSI) as a continuous measure to integrate with remotely sensed data. One to nine papers were published each year (μ = 4.1, σ = 2.2; Supplementary Fig. S2a), with no significant linear trend over time (t = 1.422, P = 0.179). The 62 articles were in 17 scientific journals; half (31) were in either Remote Sensing of Environment or International Journal of Wildland Fire (Supplementary Fig. S2b).

For the 62 studies, 352 of 401 fires were uniquely identified and geospatially located (Fig. 3a). Study fires were most commonly located in temperate conifer forests (185 fires). Less commonly studied ecosystems included boreal forests/taiga (51), temperate broadleaf and mixed forests (47), deserts and xeric shrublands (38), mediterranean forests, woodlands, and scrubs (19), and tundra (9). One fire was studied in each of the following: flooded grasslands and savannas, temperate grasslands, savannas and shrublands; and tropical and subtropical grasslands, savannas and shrublands. Fires sampled were mostly in the Northern Hemisphere, especially in the United States (283) and Canada (22). Fires were also sampled in China (17), Spain (15), Russia (5), Australia (2), Greece (2), and Belgium, Portugal, and Burkina Faso (1 each). Two fires from Miller et al. (2009) and 36 from Parks et al. (2019) lacked sufficient information to identify the locations of fires. In the United States, fires were mainly in temperate conifer forests, deserts and xeric shrublands in the West.

Fig. 3.  (a) Location of fires (black dots) and ecosystem type (N = 352 fires). The number of fires in each ecosystem type are shown in parentheses in the key. (b) Years of fire ignition (N = 401 fires). (c) Fire sizes (log₁₀ hectares) (N = 328 fires). Note: 401 fires were identified in the studies reviewed, but some lacked information on size or location.

Each study reported 1–263 fires (μ = 9.9 ± 4.9; η = 3); 23 investigated a single fire. Of the 62 studies reviewed, only 53 explicitly stated the number of fires in the analysis. The other nine studies either reported multiple or several fires (7), an unknown number (1), or did not specify (1).

The earliest identified fire occurred in 1994 and the latest in 2017 (Fig. 3b); 401 unique fires with known ignition dates were identified. Most fires (194/total) were either explicitly labelled as wildfires or classified as such by Monitoring Trends in Burn Severity, a program that maps burn severity and perimeters for fires over 1000 acres in the western USA, and 500 acres in the eastern USA (Eidenshink et al. 2007). Others were prescribed fires (74), wildland use fires (28), and a combination of prescribed and wildland fires (1). Of the remaining 104 fires, four specified only that they were caused by lightning; the rest (96) did not explicitly identify fire type, but most were likely also wildland fires.

For the 328 fires where information was available, the range of fire sizes was 28–730 855 ha (μ = 18 005, σ = 58 067; Fig. 3c); the majority (65%) were <5000 ha. The sizes of 74 fires were not explicitly given or found in supplementary searches; sometimes this information was not stated (44), was provided as a sum across several fires (3), or the study reported only the regional burn area and not the specific fire studied (1).

van Wagtendonk et al. (2004) was the first study to use CBI as a continuous measure of burn severity, following its introduction at a Joint Fire Science Conference and Workshop by Key and Benson (1999) (Fig. 4a). Studies using the GeoCBI as a continuous variable to link to remotely sensed data were first published in 2009 (De Santis and Chuvieco 2009). Most of the studies reviewed relied on these protocols, but four investigated modifications such as weighted versions (WCBI) (Soverel et al. 2010; Cansler and McKenzie 2012; Mallinis et al. 2018) or BSI (Loboda et al. 2013) (Table 3). The GeoCBI, WCBI, and BSI emerged as alternate approaches to field sampling, but CBI remained the most common composite severity measure throughout this review.

Fig. 4.  (a) Number of reviewed studies that included measurements of Composite Burn Index (CBI), Geometrically Structured CBI (GeoCBI), Weighted CBI (WCBI), or Burn Severity Index (BSI) by year (N = 62). The same study could be in multiple bins if it measured more than one composite severity index. (b) Composite severity indices used by individual studies (N = 62). The number of studies that used each index or combination of indices is in parentheses.

Table 3.  Types of field data used in studies.

Most (57) studies used a single method (e.g. CBI, GeoCBI, or WCBI) to measure burn severity on the ground; CBI on its own comprised 45 of the 62 papers reviewed in this study, and GeoCBI alone was used in 12 studies (Fig. 4b). Only three papers incorporated multiple methods of measuring severity on the same fire and were compared whether different field methods resulted in better or worse relationships with remotely sensed data (De Santis and Chuvieco 2009; Cansler and McKenzie 2012; Mallinis et al. 2018). A fourth study (Soverel et al. 2010) collected two types of field data but did not compare their performance in the analysis.

The established CBI protocol was sometimes modified for individual study purposes (Table 4). We identified five ways the method was adapted, including omitting or adding ecosystem components, modifying the strata height thresholds relative to site-specific vegetation characteristics, modifying strata weights when calculating overall plot severity, and implementing altered protocols that streamline measurements.

Table 4.  Modifications of the Composite Burn Index (CBI) protocol.

Studies varied in whether they measured unburned plots in addition to burned plots, and if they did, varied in how they included them in their analyses. Including unburned plots would extend the range of the remote sensing measurements to include values for both vegetation that did not burn and vegetation that did burn. Most commonly, the reviewed studies did not specify whether they measured unburned plots in addition to burned plots (24 studies) (Fig. 5a). However, nearly as many studies did include unburned field plots on the sampled fires (21). Additionally, several studies that sampled multiple fires included at least one fire with unburned field plots and one without (3). Two studies collected unburned field plots but excluded them from some analyses; of those, one study excluded the unburned plots altogether, and one excluded them from the continuous severity analysis but not the classification analysis. When unburned field plots were not measured on the ground, some studies incorporated ‘pseudo-unburned’ field plots (5), where areas outside the fire perimeter were assumed to have a CBI (or other burn severity measurement) score of ‘0’ or unburned. Otherwise, unburned field plots were not sampled at all, either on the ground or through remotely sensed data (6).

Fig. 5.  (a) Proportions of studies that included unburned plots (N = 62 studies). The number of studies (N) of each unburned type are in parentheses. (b) Proportions of field plots by severity class across all fires where information was (N = 21 studies). The number of plots (N) by severity class are in parentheses (N = 2968 plots). (c) Mean distribution of plots by severity class for each fire; error bars show standard errors of the means (N = 2968 plots). UB, unburned; L, low; M, moderate; H, high severity.

The distribution of field plots classified by burn severity (unburned, low, moderate, high) was provided in 21 studies (a combined 2968 plots). (1) The counts of plots were summed by severity category across studies. In this case, high severity was the most common classification (Fig. 5b); of all field data measured across all studies, there were more high severity plots than any other category. (2) Data were assessed by calculating the proportion of plots placed into different burn severity classes by each study. Here, the highest proportion of plots were in moderate severity areas, followed by high severity, then low severity (Fig. 5c). This means that each study, on average, measured more moderate-severity areas. However, the specific thresholds considered for each severity category varied by study (Table 5).

Table 5.  Composite Burn Index (CBI) minimum scores for low-, moderate-, and high-severity classes across studies that provided categorical severity thresholds (N = 17 studies).

For the 357 fires that had recorded plot counts, the average number of plots used per fire was 37.0 (σ = 53.4); 44 fires had no recorded plot counts (Fig. 6a). Both fire size and plot count information were available for 320 fires. The number of plots used per fire increased with fire size, but, as fires get more massive, the plots used increased more slowly (Fig. 6b, log-log slope estimate = 0.2246, P < 0.001, R² = 0.10).

Fig. 6.  (a) Plots counts per fire (N = 357 fires). (b) Log–log plot of number of plots used versus fire size in hectares (N = 320 fires). The grey line shows a linear model fitted through the log–log transformed data.

For the 159 fires where both the date of fire occurrence and timing of field plot measurements were reported, field campaigns occurred 1–40 months after the fire began, with an average delay of 10.5 (σ = 6.0) months (Fig. 7a). Plot size was reported for 137 fires (range = 13–6362 m², μ = 1331.0 m², σ = 1734.5; (Fig. 7b). For these fires, plots were mostly circular (105 fires) and sometimes square (32 fires).

Fig. 7.  (a) Delay of field measurements after fire ignition (N = 159 fires). (b) Field plot size (N = 137 fires).

Most studies located field plots using a stratified approach based on homogenous areas identified by dNBR, as suggested in the FIREMON Landscape Assessment protocol (Key and Benson 2006). However, two studies (Schepers et al. 2014; Warner et al. 2017) located plots in random locations. Schepers et al. (2014) used a split approach, with field measurements in both homogenous and random locations; the randomly located plots were more heterogeneous in both burn severity and vegetation type. Warner et al. (2017) used a random approach combined with relatively many samples, which enabled a simpler and more direct estimate of the accuracy of their remotely sensed severity map.

The most used sensors were from the Landsat program (Supplementary Table S6). Studies using the Landsat 4–5 (Thematic Mapper; TM) and Landsat 7 (Enhanced Thematic Mapper; ETM+) satellites were the most numerous (60 studies combined), followed by Landsat 8 Operational Land Imager (OLI; 8) and unspecified Landsat sensors (6). Most studies relied on sensors that acquired multispectral imagery with wavelengths encompassing the visible, near infrared, shortwave infrared, and thermal regions of the electromagnetic spectrum (56). Other sensors did not sample at higher wavelengths and thus did not collect shortwave infrared or thermal data. Two studies used Synthetic-aperture radar (SAR). Sensors were mounted on platforms ranging from satellites to airplanes to uncrewed aerial vehicles (UAVs).

Because the most used sensors were from the Landsat program, most studies used remotely sensed data with a 30 m spatial resolution. The coarsest spatial resolution (1000 m) was associated with the Moderate Resoltion Imaging Spectroradiometer (MODIS) on NASA Terra and Aquas satellites (Holden et al. 2010; Kolden and Rogan 2013; Hultquist et al. 2014; Zheng et al. 2016), and the finest (0.02 m) was from imagery collected by a consumer-grade RGB camera (Fraser et al. 2017).

Most studies used a single sensor (49) or included two sensors (11) (Supplementary Fig. S3a). Two studies included three or four sensors. For our analysis, all Landsat sensors (TM, ETM+, OLI) were combined, so studies using more than one sensor do not reflect those that used different Landsat satellites for pre/post fire imagery. We identified five ways that studies incorporated two or more sensors in their analysis (Table 6). All studies that used more than one sensor included data from at least one of the Landsat satellites. All but one study that used multiple sensors also collected optical and sometimes also thermal data. The exception was Tanase et al. (2015a), who included both optical and SAR data.

Table 6.  Five ways that studies incorporated two or more sensors into their analysis.

Regarding single and bitemporal remotely sensed data, studies mostly included bitemporal imagery (38) (Supplementary Fig. S3b). Sixteen studies that included bitemporal data also analysed relationships using single-date indices. The least common strategy was to include only single-date indices (8); three of these were based on Multiple Endmember Spectral Mixture Analysis.

The studies reviewed used several atmospheric correction methods for optical data; most common were top of atmosphere (or at-sensor reflectance; 23 studies) or surface reflectance derived by various methods (28) (Supplementary Fig. S4). Eight studies did not specify the level of radiometric correction. Of those, two stated only that radiometric corrections were used, and three stated only that atmospheric corrections were used. Of the studies using surface reflectance, the most commonly stated method was Dark Object Subtraction (DOS; 9), followed by the cosine of the solar zenith angle correction (COST; 4) (Chavez 1996). Other studies used unspecified surface reflectance (15) or used different methods depending on which band was assessed (1). One study quantified the effect of using four atmospheric correction methods (Fang and Yang 2014).

Some bi-temporal datasets used pseudo-invariant features (PIFs) to conduct radiometric normalisation or assess its necessity (10 studies). Most commonly, studies that conducted relative radiometric normalisation used PIFs to regress the pre-fire values against the post-fire values (4), but normalisation of post-fire reflectance to the pre-fire image also occurred (1). The remaining studies that identified PIFs found that no further relative radiometric normalisation was required (5). Except for one case, there was no mention of how PIFs were located; that particular study specified that its method used the iteratively re-weighted multivariate alteration detection (IR-MAD) algorithm to select PIFs, which were then used to normalise the target images on a band-wise basis. The other studies that did not identify PIFs stated that no scene-to-scene radiometric normalisation was applied (2) or used pre-processed imagery prepared by the Monitoring Trends in Burn Severity (MTBS), Burned Area Emergency Response (BAER), or Rapid Assessment of Vegetation Condition after Wildfire (RAVG) programs (5). The remaining 43 studies did not evaluate the need for relative radiometric normalisation.

Georeferencing/co-registration was used in some bitemporal datasets to reduce geometric error between image pairs (Table 7). Studies usually did not specify any type of georeferencing or co-registration (27 studies). For the studies that explicitly stated that they did use co-registration, the most common method was to co-register data between image pairs (10). For methods used in the remaining studies, see Table 7.

Table 7.  Specified methods for georeferencing/co-registration used by studies.

Most studies that undertook co-registration used ground control points to tie images together (specifying 30, 34, 40, 80, or enough control points to achieve specified root-mean-square error [RMSE; 5 studies]). One study used the Imagine Autosync feature in ERDAS Imagine 2015 software (Hexagon Geospatial Inc., Norcross, GA, USA), which matches images using automatically generated tie-points. Studies that reported the accuracy of co-registration (6) did so using either pixel accuracy or RMSE. Pixel accuracy was reported within 0.5 pixels (1) or 1 pixel (1); RMSE was reported using pixel units (4) at levels of 0.014, 0.16, 0.5 or lower, and 0.5. Transformations used included first-order (2), second-order (3), and third-order (1) polynomials. Resampling methods reported were nearest neighbour (2) and bilinear (1).

Most indices used in the studies reviewed were based on the optical region of the electromagnetic spectrum, which incorporates portions of the visible, infrared, and shortwave infrared regions (Fig. 8a). Less common were mixed indices, combining spectral regions with thermal indices, followed by thermal- and radar-based indices. Regarding temporal form, indices were mostly single-date or bitemporal using an absolute (pre-post) difference (Fig. 8b). Less common were relativised bitemporal indices (e.g. RdNBR, RBR). Finally, the least used types of temporal indices were bitemporal ratio indices (pre/post) (2 indices: Radar Burn Ratio, image ratioing) and combination bitemporal absolute difference and single-date index (1 index: dNBR-Enhanced Vegetation Index [EVI]).

Fig. 8.  (a) Frequencies of indices studied (N = 105 indices) according to the region of the electromagnetic spectrum used. Optical: visible, near infrared, and shortwave infrared; Mixed: combination of optical and thermal. (b) Frequencies of indices studied (N = 105 indices) according to the temporal form of the index equation. Single: single-date index; Bitemp: bitemporally differenced (pre-post) index; Relative: relativised bitemporal index; Bitemp-ratio: bitemporal ratio (pre/post) index; Bitemp-single: combination of bitemporally differenced and single-date index. (c) Frequency of indices used in at least three studies (76 additional indices were used in only one or two studies). dNBR, differenced Normalised Burn Ratio; RdNBR, Relativised differenced Normalised Burn Ratio; NBR, Normalised Burn Ratio; dNDVI, differenced Normalised Vegetation Index; NDVI, Normalised Vegetation Index; dEVI, differenced Enhanced Vegetation Index; RBR, Relativised Burn Ratio; TM5, Thematic Mapper band 5 reflectance; TM4, Thematic Mapper band 4 reflectance; TM7, Thematic Mapper band 7 reflectance; SAVI, Soil Adjusted Vegetation Index; TM3, Thematic Mapper band 3 reflectance; TM2, Thematic Mapper band 2 reflectance; TM1, Thematic Mapper band 1 reflectance; LST, Land Surface Temperature; KTW, Kauth-Thomas Wetness Transform; KTG, Kauth-Thomas Greenness Transform; KTB, Kauth-Thomas Brightness Transform; EVI, Enhanced Vegetation Index; SMI, SWIR-MIR Index; LST/EVI, Ratio of Land Surface Temperature to Enhanced Vegetation Index; dTM4, differenced Thematic Mapper band 4 reflectance; dSAVI, differenced Soil Adjusted Vegetation Index; dNDMI, differenced Normalised Difference Moisture Index; dLST, differenced Land Surface Temperature; dKTW, differenced Kauth-Thomas Wetness Transform; dKTG, differenced Kauth-Thomas Greenness Transform; dKTB, differenced Kauth-Thomas Brightness Transform; d(LST/EVI), differenced Ratio of Land Surface Temperature to Enhanced Vegetation Index. For detail on the index temporal and radiometric type as well as key reference(s), see Supplementary Table S7.

NBR and Normalised Difference Vegetation Index (NDVI) were among the most studied spectral indices. The top three indices were based on NBR: dNBR (50 studies), followed by RdNBR (25), and NBR (19) (Fig. 8c). The next most common indices were the differenced NDVI (dNDVI, 18 studies) and NDVI (12). Although 105 indices were related to CBI, most appeared in only one or two studies (76 indices; see Supplementary Table S6).

The mean number of indices used in each study was 5.0 (σ = 5.7), with up to 26 indices in a single study (Fig. 9a). Absolute difference pre/post-fire indices were the most common (50 studies), followed by relativised bi-temporal indices (26), single-date post-fire indices (23), and ratio indices (3). Most commonly, studies included only absolute indices (21, Fig. 9b). Many studies had both absolute and relativised indices (13); absolute, relativised, and single-date indices (10); or just single-date indices (8). Less commonly, studies reported a combination of single and absolute indices (4) or only relativised indices (3). Least used were absolute, single-date, and ratio indices; absolute and ratio indices; or just ratio indices (1 study each).

Fig. 9.  (a) Number of indices used per study. (b) Proportions of studies using indices of different temporal form. Number of studies for each method is shown in parentheses.

Most of the studies reviewed (54 of 62) included dNBR as a spectral index. However, the majority of those studies (34) did not include a dNBR offset value (Fig. 10). We found that eight studies did not specify whether an offset value was included or not. Six studies did explicitly state that an offset was included. Additionally, two studies provided analyses both with and without an offset, and two other studies tested whether including an offset value improved the relationship with field data but ultimately excluded it from their analyses. Finally, two studies used MTBS/BAER data, which include an offset value for RdNBR indices, but not in the calculation of dNBR.

Fig. 10.  Number of studies that included a dNBR offset for those that used dNBR as an index (N = 62 studies). The number of studies for each method is shown in parentheses.

Many studies (22) did not specify how the value of indices for field plots was determined (Table 8). The studies that did specify their method most commonly extracted the value of the pixel over plot centre (14), used a 3 × 3 focal mean window centred on the plot (12), or bilinear smoothing (7). For smoothing types used in the remaining 9 studies, see Table 8.

Table 8.  Pixel value extraction methods by citation including parameters and frequency of use in studies (N = 62 studies).

Only two studies quantitatively compared different smoothing methods; one compared bilinear with no smoothing (Stambaugh et al. 2015), and one used four methods for Quickbird imagery in their comparison with ASTER and Landsat data (Holden et al. 2010).

Fifteen classes of models were used in analyses (Table 9); the most common was linear regression (36 comparisons), followed by quadratic regression (17), and exponential models (9). For the models used in the remaining 30 comparisons, see Table 9.

Table 9.  Models used in analysis and (a) whether field data were a predictor or (b) response variable, (c) analysis was based only on correlation, or (d) not specified (N = 108 comparisons).

Field data were much more commonly used as a predictor variable (75 comparisons) than as a response variable (19). This pattern held across almost all general model types. However, for three models, field data were used solely as a response variable: the natural logarithm model, the exponential models, and the sigmoidal model. For cubic and unspecified non-linear models, the same number of comparisons used field data as a predictor variable and a response variable. Eight analyses just reported correlation statistics between the two observation methods.

Many methods were used to assess or quantify the performance of models that related remotely sensed data to field plots (Table 10). The most common metric was R² (46 studies), followed by RMSE (17), R²_adj (14), p (13), and r (13). Most studies used a single evaluation metric (24) or two metrics (27, Supplementary Fig. S5). Eleven studies assessed three or four metrics.

Table 10.  Model performance metrics and their frequency of occurrence.

The CBI rating encompasses five vegetation strata, commonly separated into understorey (A–C) and overstorey (D, E) components (Fig. 1). Of the 62 studies reviewed, 14 modelled CBI across strata. Three methods investigated whether relationships between field observations and remotely sensed data varied by strata (Table 11). The most common method separately modelled the understorey components (A–C), overstorey components (C, D), and overall severity (5 studies). Studies also modelled each individual stratum separately (4), or modelled the soil or substrate component (A) and surface or vegetation components (B–E), as well as the composite or overall severity (3 studies). Only two studies modelled just the overstorey components (D, E) and overall severity.

Table 11.  Different ways of partitioning ecosystem strata for modelling.

Most studies analysed single-fire local models (33); i.e. relationships between field plots and remotely sensed data were fire-specific (Supplementary Fig. S6). Other studies investigated regional models (19), combined field plots across multiple fires, or evaluated both local and regional models (10).

This review highlights the many ways that continuous measures of severity, based on the CBI, have been linked to remotely sensed data, and outlines the range of methods and decisions in the process. Five main topics were investigated: (1) study information; (2) fire data; (3) field data; (4) remotely sensed data; and (5) linking models. Overall, we found wide variability in the methodological decisions and analytical tools used by the included studies. This finding raises the challenge of knowing whether differences among study results are driven by site-specific ecological and fire differences or were affected by differences in the methodological approaches. One outcome of this study was to provide flowcharts that highlight decisions that can be made during field data collection, remotely sensed data collection, and modelling phases (Figs 11–13). These flowcharts may facilitate more comparative studies that assess the robustness of the workflow to differences in methodological choices and highlight potential future research directions. Combined with the breakdown of methods and number of studies that have investigated each decision, future studies may focus on specific areas of the analytical framework that are not well represented (such as the effectiveness of thermal, radar, and lidar data compared with spectral imagery), or they may seek to conduct more extensive meta-analyses comparing multiple methodological choices. Additionally, we found several key research gaps that warrant future investigation. Finally, we encourage future studies to provide explicit description and justification behind their methodological approaches to make comparisons among findings in separate studies more possible.

Fig. 11.  Analytical decisions to be made during field data sampling.

Fig. 12.  Analytical decisions to be made during remotely sensed data acquisition and processing.

Fig. 13.  Analytical decisions to be made during modelling phase.

There is no one consensus for the best way to connect field observations to remotely sensed data, and uncertainties exist at every decision point. First of all, there is no universal, one-size-fits-all approach to evaluating burn severity (Cocke et al. 2005; Keeley 2009). Although some studies suggest more utility, and that tree-level measurements that directly assess ecological effects may be preferable to a composite index that aggregates semi-quantitative estimates of change across a plot (Morgan et al. 2014; Furniss et al. 2020), the choice of assessment should be determined by management, ecological purposes, and field sampling designs (Ryan and Noste 1985). We identified, where available, analyses that compared important decision points in linking field observations of CBI as a continuous variable to remotely sensed data. The numerous analytical decisions and paths that studies applied prevented strong generalisations from being made, especially considering that relatively few studies compared multiple approaches at key decision points in the workflow. Of those that did, many returned mixed results. Consequentially, the choice of how burn severity is measured and modelled influences study outcomes.

One such example relates to the comparison of different composite severity measures (CBI, GeoCBI, and WCBI). Three studies (De Santis and Chuvieco 2009; Cansler and McKenzie 2012; Mallinis et al. 2018) compared multiple ground measures of severity and found that no one method seemed to result in the greatest correspondence between field- and satellite-based estimates of burn severity across different regions. For both Cansler and McKenzie (2012) and Mallinis et al. (2018), who both included analysis of CBI, GeoCBI, and WCBI, the best performing field measurement depended on which spectral index (and in the case of Mallinis et al. (2018), which sensor) was used. De Santis and Chuvieco (2009) compared CBI and GeoCBI, and found that GeoCBI peformed best for two of the three fires assessed, and CBI performed best on the third. Thus, although metrics of fractional cover, leaf area index, or other weighting approaches did in some cases improve burn severity models, the mixed results failed to provide a convincing answer as to which composite severity measure is best.

Another example of the lack of consensus about how to sample fires relates to the inclusion of unburned and low-severity field plots in the field sample set, which provides an anchor at the low range of CBI for assessing relationships with remotely sensed data. The bias that we identified of field plots tending towards moderate- and high-severity areas likely occurs for two reasons: (1) attempts to overcome the saturation of remotely sensed indices at high severities (van Wagtendonk et al. 2004; De Santis et al. 2010; Veraverbeke et al. 2012; Parks et al. 2014); and (2) the importance of these areas to management (Robichaud 2000; Miller and Thode 2007; French et al. 2008; Cansler and McKenzie 2012; Fernández-García et al. 2018b). We found substantial variation in the thresholds used in classification, but it was clear that the underrepresentation of lower range of severities occurred despite that variation. The fact that only one of the two studies that excluded unburned plots from their analysis (Murphy et al. 2008) provided a rationale (the unburned sites (CBI = 0) disproportionately influenced the linear fit to produce remaining non-liner residual structure) highlights the need for more comparative analyses in this realm.

These examples highlight the uncertainty that often remains following quantitative assessments of methodological approaches. This is not unique to methods based on composite severity indices, however – Furniss et al. (2020) found that estimates of tree-level metrics such as mortality of stems and basal area could vary widely for a single spectral index value. Such uncertainties are unsatisfying and provide reason for caution when applying methodologies developed in one context, for example based on a single fire or ecosystem type, outside the scope of inference. For example, some influential studies (like Cansler and McKenzie’s (2012) comparison of multiple pixel value extraction methods) are widely cited by subsequent investigators but without validating their effectiveness in subsequent studies or in new regions. However, such studies may be highly site- and context-specific, and it is unclear how broadly applicable their results are. In other cases, attempts to compare different approaches, such as using sensors with varying spatial resolution, are confounded by lurking variables – in this case, the spectral resolution of different sensors and the spatial synchrony with field plots (van Wagtendonk et al. 2004; Holden and Evans 2010; Tanase et al. 2015a; Mallinis et al. 2018; García-Llamas et al. 2019). While several of these studies found slight improvements for sensors with high spatial resolution compared with Landsat imagery (Holden and Evans 2010; Mallinis et al. 2018; García-Llamas et al. 2019), it is difficult to disentangle their results to isolate the effects of the spatial resolution of the sensors. Finally, there may not yet exist any comparative studies in the current literature assessing the effect of georeferencing remotely sensed data on model results. Although 27 of 62 studies conducted or investigated the need for georeferencing (or co-registration between image pairs), no studies presented comparative analysis of the impacts on model results.

Realistically, only studies that replicate their methodological approach (e.g. use similar methods of collecting field observations, include unburned field plots, extract pixel values over plot centre with the same method, etc.) can be compared across each other. It is not well-quantified what level of uncertainty arises in comparing results across different studies where methodologies were not consistent. Therefore, more broad-scale comparative studies are needed to understand the transferability of methods across landscapes and methods/technologies.

Studies have investigated individual slices of the ‘decision menu’ (i.e. the many potential analytical approaches to a particular step in the analysis; Fig. 2), but none have looked at how uncertainty compounds across the whole analysis process from beginning to end. Stambaugh et al. (2015) assessed the compounding uncertainty across three decision points: (1) two index value extraction methods (none and bilinear interpolation); (2) two timings of remotely sensed data (initial vs extended assessment); and (3) two spectral indices (dNBR vs RdNBR). Their analysis thus followed 2³ = 8 different pathways and showed that R² for models of overall CBI values ranged from 0.03 to 0.61. This study demonstrates the importance of understanding not just how each slice of the decision menu can affect model results but also how those uncertainties can be compounded through different analytical pathways. Additionally, because many of the comparative studies discussed above provide mixed results when looking at a single decision point, it is unclear how important such choices are at the global scale of analysis. A comprehensive study of the many potential analytical approaches would help understand when differences at different decision points wash out and when they do not, allowing us to understand when we are seeing ‘true’ differences among studies and when those are due to differences in the methods used.

We identified several key research gaps or biases in the studies reviewed. The previous sections addressed the importance of comparative studies targeting key points in the ‘decision menu’ of an analysis. Here, we identify where individual studies are absent or lacking from the literature identified in our review.

The uneven distribution in the types and locations of fires studied has important implications for how widely applicable the results are across different fires and geographies in the future. First, most fires were wildfires, with some prescribed and wildland fire use fires. Although this makes sense from a perspective of pre-fire management and post-fire intervention, there is reason to question whether prescribed fires behave in fundamentally different ways than wildfire. For example, Arkle and Pilliod (2010) found that prescribed fires did not mimic the ecological effects that followed wildfire in a riparian ecosystem; both the extent and severity of vegetation burned was substantially lower in the prescribed fire compared with nearby wildfires. How severity models based on remotely sensed data under such conditions might differ remains unclear. Some landscapes, such as those in the southeast USA, experience primarily prescribed fires – where resources for post-fire assessment may be more heavily allocated to determining if the burn objectives were successfully achieved. In those cases, the CBI protocol may be overly broad, and other, more direct measures of ecosystem change may be preferred (Morgan et al. 2014).

Second, there was a heavy geographical bias towards the USA and especially western USA, as well as a heavy focus on temperate coniferous forests. These results reflect that western USA (1) has a long history of wildfire research; (2) experiences large fires in many forested areas; and (3) has high spatial variability in fire regimes and biodiversity (Heyerdahl et al. 2001; Morgan et al. 2001). However, the concentration of fire research in this area may also mean that major swathes of ecosystems elsewhere are unstudied. Non-forested areas, such as frequently burning grasslands, aren’t suitable for assessments with CBI (i.e. it is difficult to classify burn severity as anything other than burned and unburned). However, some regions that are forested and experience frequent fires or have extensive fire histories (e.g. Italy, France, Indonesia, or Brazil) are not highly represented by CBI studies. One reason for this could be different research traditions or regional research history. For example, in Australia, alternative visual estimation approaches have been used (Hammill and Bradstock 2006; Chafer 2008) instead of CBI. A second reason may be due to differences in the underlying ecology of post-fire responses. For example, some forests grow back in very different ways (e.g. resprouting eucalypts) compared with temperate and boreal forests, where CBI was originally developed. Although other regions may have different ways of conducting field estimates of severity not included in this review, the ability to make comparisons between those studies and CBI is limited.

The limited range in the size and timing of field plot collection limits our understanding of how field data at different scales affects model variability and of long-term ecosystem change. First, there was a strong bias towards 30-m diameter plots, raising the question of what spatial scales are missing in current analyses. The CBI was initially designed with 30-m diameter plots to match the spatial grain of imagery collected by the Landsat satellite program (Eidenshink et al. 2007), but has since been adapted and related to many different sensors of both higher and lower resolution (Supplementary Table S6). Second, we found a relative lack of more extended intervals between fire occurrence and field data collection (on the order of 3–10 years post-fire), as well as an absence of studies conducting repeated measurements.

Although the collection of field data is often limited by site accessibility, high costs, and time constraints (Chuvieco et al. 2006; Boucher et al. 2017), there are clearly opportunities for studies that sample outside the range of commonly used field plot sizes and timing delays. For example, studies could collect very large field plots (e.g. 250 m diameter); however, such an endeavour might take weeks to months, and environmental factors such as rain or wind could affect measurements (Meng and Meentemeyer 2011). It remains unclear how sampling very large field plots may impact relationships with remotely sensed data, especially those of relatively coarse spatial resolution. Additionally, studies could collect plots at extended times post-fire. However, it becomes more difficult to see burn severity effects on the ground. Most studies conform to the 1–2 year time frame of ‘initial’ and ‘extended’ assessments described by Key and Benson (2006), and the effectiveness of the CBI protocols over extended time-periods has not been well studied. This could be because burned areas can become very hazardous to assess because of falling trees and branches, making access challenging over time.

The 15 model forms used across the studies begs the question of how robust results are to uncertainty in model form. The sheer number of model forms in the studies reviewed shows the diversity of statistical approaches used to link remotely sensed data to field observations. We identified two main considerations in the choice of model form: (1) the behaviour of remote sensor and ecosystem considerations (sensor saturation, nonlinearity at high severities); and (2) model performance (based on evaluation metrics). Although studies such as van Wagtendonk et al. (2004) modelled CBI using a quadratic model, the predicted reduction in CBI at high dNBR values is not an expected behaviour of the sensor-ecosystem dynamic (Hall et al. 2008). Thus, some studies argued for nonlinear models with saturated growth characteristics (e.g. CBI = dNBR × (a[dNBR] + b)⁻¹; Hall et al. 2008), even when this choice is not strictly supported by model evaluation metrics. The motivations for selecting different model forms were seldom described, but many studies clearly tried to find the best descriptor of the observed relationship. In studies that considered the sensor-ecosystem dynamic, predicting a trend that is feasible in the real world overrode the aim for strong performance (Hall et al. 2008; Boucher et al. 2017). That said, studies rarely compared more than a few models, and the robustness of relationships to uncertainty in model form is not well understood.

Finally, the mixed results in the 14 of 62 studies that modelled CBI across strata demonstrated that conceptual models of remote sensed data reported strengths or weaknesses in correlating with specific burn severity measures that would not be expected from theory. Specifically, we identified some studies showed stronger relationships in lower canopy strata – in conflict with accepted theory that passive remotely sensed data are limited by low signal penetration (Chen et al. 2015) and disproportionately capture changes in the upper canopy (Kasischke et al. 2008). Several studies support the theory that passive sensors are more likely to detect fire effects in the upper part of the vegetation stratum (Patterson and Yool 1998; Hudak et al. 2004; De Santis and Chuvieco 2009). Conceptually, the ability to capture change in lower canopy strata is impacted by vegetation density, because burned landscapes may reflect more changes in the overstorey strata due to the shielding effect of vegetation or their remains (Soverel et al. 2011; Tanase et al. 2011). Although relationships between field observations and remotely sensed data were generally strong in the overstorey components, large tree stratum, and vegetation components (Chen et al. 2011, 2015; Meng and Meentemeyer 2011; Stambaugh et al. 2015; Wu et al. 2015; Warner et al. 2017), two studies found that the understorey components outperformed higher canopy strata (Hoy et al. 2008; Tanase et al. 2011). These exceptions to the conceptual rule-of-thumb that passive remotely sensed data are limited in their ability to detect change in lower canopy strata deserve more scrutiny and better understanding of the conditions under which lower forest strata may be observed.

Our research highlighted the importance – especially given the lack of consensus on analytical methods – of providing rationales for each decision in the analysis workflow that link back to the research objectives. For example, in regards to the two previously stated considerations that drove the choice of model form in the reviewed studies, it would be beneficial for future investigators to more explicitly give the rationale behind their model selection methods and state the conditions under which those models can be appropriately interpreted. Choosing to sacrifice model performance for predictions that correspond to expected behaviour of the sensor and ecosystem dynamic may create difficult comparisons across fires or regions, but it also provides considerable benefit in that they reflect the true nature of the system. Both approaches are equally valid given conforming research objectives, but we suggest that studies (1) consider the purpose of their analysis when determining appropriate models to assess and (2) convey whether weaker model performance metrics may result from their decisions.

Similarly, we found that studies did not always state their rationale for using field data as either predictor or response variables depending on their research objectives. In typical regression analysis, a causal relationship is implied between one or more predictors, and a response (Bordacconi and Larsen 2014). Cansler and McKenzie (2012) used this logic to argue for CBI as a predictor variable, because burn severity changes reflectance, not the other way around. They further stated that CBI has the greatest certainty associated with its meaning and thus should be used to predict the variables with no inherent ecological meaning (e.g. dNBR or RdNBR). Conversely, in some cases CBI (or other ground severity measure) may serve as a response variable, with remotely sensed data serving as the predictor variable if the main study goal is to predict severity on the ground (Zhu et al. 2006). In these cases, studies taking this approach follow the logic that the known value in a satellite image (e.g. RdNBR) is the predictor variable being used to model the unknown value on the ground (e.g. tree mortality) as the response variable. Both approaches have their benefits and challenges, though readers could benefit from more studies explicitly describing when and why field data are considered as predictor versus response variables, because model configuration affects the ability to compare across studies.

The analytical process of modelling of continuous composite severity measures using remotely sensed data can be split into three main phases: (1) collection of field data; (2) collection of remotely sensed data; and (3) modelling. We provide figures that break down the main decisions that can be made at each phase (Figs 11–13). This blueprint for preparation and analysis should be useful for considering key decisions in study design as well as emphasising the choices that should be justified when reporting study results (see example rationales in Table 12).

Table 12.  Key decisions and potential rationales for each phase of study design and analysis.

Many decisions discussed in this paper revolve around image quality and availability, and problems with comparability and synthesis across disparate collections of CBI data result from this limitation. The selection of suitable remotely sensed imagery requires consideration of many different characteristics, including solar zenith angle, atmospheric effects, plant phenology, and the availability of cloud-free imagery (Rogan et al. 2002; Epting et al. 2005; De Santis and Chuvieco 2007; Ju and Roy 2008). If bi-temporal indices are used, then inter-annual meteorological differences (Veraverbeke et al. 2010) and the misregistration of image pixels (Verbyla and Boles 2000) must also be considered. Sensor differences and some atmospheric effects may be corrected through radiometric normalisation methods, but differences in vegetation phenology may be unavoidable, particularly in cloudy regions where there may be few cloud-free images available (Epting et al. 2005). Furthermore, the seasonality and lag timing of post-fire imagery can impact the values of spectral indices. For example, RdNBR is sensitive to ash cover, which declines with time since fire (Miller and Quayle 2015). Therefore, RdNBR values that represent total mortality can be different immediately post-fire compared with the following year. Veraverbeke et al. (2010) provide an in-depth discussion of the effect of seasonal timing on index values.

Although limitations in image availability interact and potentially exacerbate the other problems/limitations that were identified in previous studies, our review focuses on the wider set of methodological practices. Future meta-analyses could focus on CBI datasets that cover broader extents (Picotte et al. 2019) and spectral index processing methods that are less reliant on individual image selection techniques (Parks et al. 2021). Studies using these additional datasets and approaches will likely impact accepted research practices and potentially lead to more consistent methodologies. Currently, researchers may want to consider the effects of those limitations along with other issues we raise in this paper.

One major limitation of this study is that we did not evaluate the specific equations used to calculate each spectral index included in the studies. A few references provide incomplete equations – for example, in the original Miller and Thode (2007) study presenting RdNBR, they do not include multiplying by 1000 explicitly in the equation, nor any modifications needed to deal with 0 in the denominator. Additionally, many studies do not explicitly include the use of offset values in the equation. We captured this information where available but recommend that authors more explicitly state the form of the equations for spectral indices as well as any use of offset values. We did not endeavour to include the correct equations here (Supplementary Table S7), but we note the issue with past inconsistencies.

Many of the North American studies used overlapping CBI datasets (so the datasets are not really independent – e.g. Cansler and McKenzie (2012) data were used by Karau et al. (2014) as well as Parks et al. (Parks et al. 2014, 2018, 2019). We also note that much of the CBI datasets analysed by the papers in this review are available in a data repository for reanalysis (Picotte et al. 2019).

We did not include studies in our review that used direct forest measurements, for example the percentage of tree mortality or char and ash colour (see Table 1). This may have introduced some geographic bias in which studies were excluded because some regions may rely more heavily on severity measurements based on direct, individual metrics. For example, Morgan et al. (2014) recommend recording actual fire measurements that have logical and mechanistic connections to the properties a sensor can detect, avoiding the use of composite measures such as CBI that collapse multiple ecosystem attributes into a single index. Several studies have begun to record direct measures of burn severity rather than CBI only (Whitman et al. 2018; Harvey et al. 2019; Saberi et al. 2022), though detailed field measures of burn severity remain rare compared with the widespread use of CBI. Another alternative approach to characterising burn severity over large areas uses lidar data to more directly measure changes in forest structure (McCarley et al. 2017). Although using multi-temporal lidar to examine changes in forest structure due to fire would allow more precise measurement of individual metrics (e.g. change in canopy cover), reduce observer bias, and allow quantification of change in areas without nearby unburned reference sites, the approach remains limited by the availability of pre- and post-fire lidar at adequate time intervals.

Given the difficulty of synthesising results across studies that used varying methodologies, we support efforts such as Picotte et al. (2019) to aggregate and disseminate datasets based on composite severity field observations from investigators. This could provide the necessary information to conduct a meta-analysis of the effects of compounding error across the ‘decision menu’ presented in this review. Additionally, we recommend future studies complete the following:

1. Collect and share field data in such a way that the original and modified composite severity measures identified in this review (CBI and GeoCBI/WCBI) could all be calculated, thus facilitating further comparisons among the different field measures. Such a request may require the development of a new CBI field collection form for collection of raw field measurements that could be used to calculate any of these.

2. Report full detail of any processing and analysis. For this review, we followed up with authors where information was missing; however, we were not able to obtain information requested in every case. For a list of recommended information to report, Table 13.

3. Report the reasoning for methodological choices at each step in the analysis. One of the goals of this review was to identify key analytical decisions and potential reasons why investigators may want to make one decision vs another (Table 12).

Table 13.  Key elements suggested to report for each phase of study design and analysis.

This review of studies linking remotely sensed data to continuous measures of burn severity measured with the Composite Burn Index highlights: (1) the wide range in analytical approaches and lack of consensus in methodological decisions; (2) the scarcity of comparative studies at any one point in the ‘decision menu’ and absence of a comprehensive beginning-to-end quantitative analysis of compounding uncertainty throughout the analysis framework; and (3) key gaps in research relating to the distribution in the types and locations of fires studied, limited range in the size and timing of field plot collection, and modelling of CBI across strata.

The results of this review provide a framework for future studies linking remotely sensed data to continuous measures of severity by summarising the key analytical decisions and the distribution of studies using such techniques. We avoid concluding which methods or decisions perform best and instead focus on the importance of understanding the wide varieties of ways this type of research has been done, and its potential impacts on our understanding of the state of the science. We find that much uncertainty remains in light of a lack of comparative analysis and biases in the study designs. In the absence of a consensus approach to modelling severity using remotely sensed data, we suggest future research explicitly state their rationales for each analytical decision and how it relates to their specific research questions.

Supplementary material is available online.

The data that support this study will be shared upon reasonable request to the corresponding author.

The authors declare no conflicts of interest.

This work was partially supported by the USDA Forest Service Pacific Northwest Research Station [Agreement Numbers: 17-PA-11261987-043, 20-JV-11261987-035], and the Precision Forestry Cooperative, University of Washington, Seattle, Washington, USA.

Conceptualisation, C. W. M. and B. J. H.; data curation, C. W. M.; formal analysis, C. W. M., funding acquisition, C. W. M, E. A., and L. M. M.; investigation, C. W. M. and B. J. H.; project administration, C. W. M.; resources, C. W. M., E. A., and L. M. M.; software, C. W. M.; supervision, B. J. H., V. R. K, E. A., and L. M. M.; validation, C. W. M.; visualisation, C. W. M.; writing – original draft, C. W. M., writing – review and editing, C. W. M., B. J. H., V. R. K., E. A., and L. M. M. All authors have read and agreed to the published version of the manuscript.