Soil Burn Severity Assessment Using Sentinel-2 and Radiometric Measurements

Abstract

The objective of this article is to create soil burn severity maps to serve as field support for erosion tasks after forest fire occurrence in Spain (2017–2022). The Analytical Spectral Device (ASD) FieldSpec and the CIMEL CE-312 radiometers (optical and thermal, respectively) were used as input data to establish relationships between soil burn severity and reflectance or emissivity, respectively. Spectral indices related to popular forest fire studies and soil assessment were calculated by Sentinel-2 convolved reflectance. All the spectral indices that achieve the separability index algorithm (SI) were validated using specificity, sensitivity, accuracy (ACC), balanced accuracy (BACC), F1-score (F1), and Cohen’s kappa index (k), with 503 field plots. The results displayed the highest overall accuracy results using the Iron Oxide ratio (IOR) index: ACC = 0.71, BACC = 0.76, F1 = 0.63 and k = 0.50, respectively. In addition, IOR was the only spectral index with an acceptable k value (k = 0.50). It is demonstrated that, together with NIR and SWIR spectral bands, the use of blue spectral band reduces atmospheric interferences and improves the accuracy of soil burn severity mapping. The maps obtained in this study could be highly valuable to forest agents for soil erosion restoration tasks.

1. Introduction

Forest fires are one of the most common causes of ecosystem disturbance, such as vegetation loss, soil erosion, and runoff [1,2]. In countries such as Spain where summers are becoming hotter and drier, forest fires occur more frequently and are larger in size, causing environmental and economic losses [3].

One of the major problems encountered at the environmental level is the risk of erosion, which is closely related to the soil burn severity after fires [4,5,6,7]. Soil burn severity is defined as the level of change caused by fire at the level of organic cover, biomass burning, and mineral soil [8,9]. Both in Atlantic climates (where high levels of precipitation influence vegetation development by preventing soil monitoring) and in Mediterranean climates (where the abandonment of rural areas and the increase in temperature due to climate change led to an increase in accumulated fuel), it is considered of vital importance to assess the risk of erosion in order to prevent greater damages (increased probability of flooding, loss of flora and fauna, …) [10,11,12,13,14]. In this sense, a correct and accurate prediction of soil burn severity associated with erosion risk is necessary, although challenging [15].

The use of remote sensing is a reliable and accurate technique for the study of forest fires [16]. Satellite sensors make it possible to monitor changes quickly, easily, and remotely [17,18]. Historically, several sensors on board earth-orbiting satellites and aircrafts have been used to observe and detect forest fires [9]. One of the most widely used sensors for the study of forest fires today is the MultiSpectral Instrument (MSI) on board the Sentinel-2 satellite [19]. Due to its spatial (10–60 m) and spectral resolution (443–2190 nm), several authors recommend its use for analysis of burned areas and fire severity (vegetation + soil) [20,21,22,23,24,25,26,27,28]. Its temporal resolution (3–5 days), allows us to analyze changes in a burned area in a continuous way, except in areas with a high presence of clouds [23,29]. The spectral behavior of Sentinel-2 bands, such as near-infrared (NIR) and shortwave infrared (SWIR) spectral regions, are relevant for detecting changes in forest cover [30] and changes in landscape moisture [31], respectively. In pre-fire conditions, the NIR region shows high reflectance values related to healthy vegetation. Conversely, the SWIR region shows low reflectance due to water absorption [32]. After post-fire, the NIR reflectance decreases due to a leaf pigment and cell structure reduction [33]. In contrast, SWIR reflectance increases due to the drying effect of the fire [34]. Red-Edge spectral region allows monitoring chlorophyll content absorption in vegetation [29]. In addition, synergy between blue and red spectral regions minimizes the effects of atmospheric scattering caused by aerosols in the red channel [35].

Together with optical sensors, the thermal spectrum is also a useful source of information in forest fires applications [36,37,38,39,40]. In particular, the emissivity parameter, defined as the ratio between the object emitting capacity and that of a blackbody at the same temperature, shows changes when soil properties, such as moisture content, organic matter, and mineral composition, are affected by forest fires [41]. In contrast to temperature, emissivity remains constant allowing a homogeneous classification of the study area without depending on image acquisition time. Sensors such as the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) and the Moderate Resolution Imaging Spectroradiometer (MODIS), onboard TERRA satellite and the Thermal Infrared Sensor (TIRS), and onboard Landsat-8 satellite are commonly used in multiple forest fires applications [36,37,38,39,40].

Although remote sensing is widely used in the forest fires field, there are several restrictions that impede soil burn severity estimation with the same reliability as fire severity [42]. These restrictions are due to factors such as interference between atmospheric absorption and soil absorption and the interference of burned trees shadows. Moreover, the presence of different soil burn severity levels on the field is much more heterogeneous (one pixel contains a mixture of spectral response of different canopies) compared to different fire severity levels focused on vegetation strata, where large areas of land sharing the same fire severity can be observed [43].

As a method to support remote sensing, techniques such as optical radiometers (350–2500 nm), which have proven to be useful in the study of soil properties within a forest fire [12,27], can be used. The Analytical Spectral Device (ASD) FieldSpec radiometer was used in several studies related to forest fires [44,45,46]. In addition, thermal radiometers are also useful in forest fires assessment. In this sense, CIMEL CE-312 radiometer (8–14 µm) enables the precise measurement of brightness temperature from which emissivity can be obtained for 6 different channels. In contrast to satellite or airborne sensors, optical and thermal radiometers allow repeated in situ measurements with very high spatial resolution (depending on the FOV), spectral resolution (hyperspectral sensor with a bandwidth of less than 10 nm), and minimize the atmospheric effect on the measurements.

However, for both satellite-borne sensors and radiometers, in order to choose the best areas to measure and validate the results obtained, it is important to obtain a classification that describes the soil burn severity. This classification allows an easy identification of different soil burn severity levels from visual indicators (e.g., changes in color or structure) [6,43]. One of the most widely used classifications for fire severity validation is the Composite Burned Index (CBI), which visually estimates the impact of fire on combined vegetation and soil strata. However, several authors advise the need to split the assessment of fire severity into vegetation and soil, pointing out the different fire behavior for both canopies and for different vegetation types [47,48,49,50]). Other widely used methods, such as the classification proposed by [6], focus the classification exclusively on the damage suffered by the ground, focusing on those areas where the vegetation has been completely consumed (high fire severity).

The aim of this article is to create soil burn severity maps to serve as field support for forest agents in charge of eroded soil rehabilitation after forest fire occurrence in Spain (2017–2022). Due to the heterogeneity of soil burn severity levels found on small surfaces (less than pixel size) and the few methods that focus on the study of soil burn severity (non-vegetation), the main challenges encountered throughout this study are the following: (1) Determining the spectral region that best discriminates between soil burn severity levels; (2) Analyzing the correlation between radiometric measurements and reflectance obtained with satellite sensors; (3) Choosing the classification map that obtains the most accurate statistical results.

2. Materials and Methods

2.1. Study Area

This study is focused on large forest fires (>500 Ha) that occurred in Spain between 2017 and 2022 (Figure 1). Forest fires studied involve two climatic regions: Mediterranean and Atlantic [51,52]. The Mediterranean region has mean temperatures and rainfall round 3–10 °C and 100 mm, respectively, and hot summers (mean temperatures and rainfall round 20–25 °C and 25–50 mm, respectively). The Atlantic climate has abundant rainfall (>750 mm) and an average temperature between 9 and 15 °C during the whole year. The mean temperature (9–11 °C) increases in the mainland hilly regions (12–15 °C). Burned areas analyzed cover over 100,000 Ha, classified mainly as Pinus sylvestris L., Pinus pinea L., Pinus pinaster Ait., Pinus radiata, Pinus nigra, Populus × canadensis, Populus nigra, Quercus pyrenaica, Quercus ilex, Quercus suber, Acacia dealbata, Alnus glutinosa, Fraxinus angustifolia, Juniperus oxycedrus, Juniperus communis, Castanea sativa, Salix spp., Olea europaea, Arbutus unedo, Eucalyptus nitens, Eucalyptus camaldulensis, and Eucalyptus globulus Labill. with a shrubby understory of Erica L. sp., Cistus L. sp., and Cytisus Desf. Sp. The burned affected soil is divided between Leptosols and Cambisols developed on slates and granite and gneiss with sandy loam texture.

2.2. Field Plots

503 sampling plots were established in areas where the vegetation was fully consumed by crown fire (minimum size 60 m × 60 m) and where stone cover (particle diameter > 10 cm) was <50%. We assessed soil burn severity in plots of radius 20 m (Sentinel−2 pixel size), centered in the larger plots. The soil burn severity was assessed with the aid of a 20 cm × 20 cm quadrat, which was placed at 80 systematically selected points along two perpendicular transects. The soil in each quadrat was classified using a modified version [6] of the soil burn severity index proposed by [43]. That classification considers six soil burn severity levels: (Level 1) very low soil burn severity: burned litter (Oi) but limited duff (Oe + Oa) consumption; (Level 2) low burn severity: Oa layer totally charred and covering mineral soil, possibly some ash; (Level 3) moderate soil burn severity: forest floor (Oi + Oe + Oa layers) completely consumed (bare soil), but soil organic matter not consumed and surface soil intact; (Level 4) high soil burn severity: forest floor completely consumed, soil organic matter in Ah horizon consumed and soil structure altered within a depth of less than 1 cm; (Level 5) very high soil burn severity: same as high soil burn severity but within a soil depth equal to or greater than 1 cm; and (Level 6) the same as Levels 4 and 5 and color altered (reddish). Due to the impossibility of detecting differences in depth, soil burn severity levels 4, 5, and 6 were grouped into a single level (high severity). None of the field plots obtained were classified as Level 1. The field campaign dates and spatial distribution of soil burn severity plots are compiled in

Table 1. Forest fire date, field campaign date, and number of soil burn severity plots for each forest fire.

Community	Place Name	Forest Fire Extinction Date	Field Campaign Date	Soil Burn Severity Plots
				Low	Moderate	High
Andalucía	Nerva	2 August 2018	4 October 2018	2	7	-
	Estepona	12 September 2021	21 October 2021 22 October 2021	2	5	-
Castilla La Mancha	Hellín	25 July 2021	18 August 2021 19 August 2021	4	7	-
Castilla y León	Real Sitio de San Ildefonso	4 August 2019	5 September 2019	4	6	2
	Pedro Bernardo	28 June 2019	2 August 2019	1	3	-
	Zamora	24 June 2022	15 July 2022	31	-	5
Galicia	Silleda	15 October 2017	14 November 2017	2	6	2
	Nigrán	15 October 2017	14 November 2017	-	1	-
	Soutomaior	15 October 2017	14 November 2017	6	17	3
	Fornelos de Montes	16 October 2017	14 November 2017	3	5	-
	As Neves	15 October 2017	14 November 2017	2	14	-
	Flariz	24 July 2020	31 July 2020 27 August 2020	13	9	1
	San Millao	29 July 2020	4 August 2020 3 September 2020 10 September 2020	12	16	1
	Verín 1	22 July 2020 4 August 2022	9 September 2020 22 August 2022 23 August 2022	12	5	-
	Lobios	12 September 2020	21 September 2020 28 September 2020	11	5	3
	Vilariño	13 September 2020	22 September 2020	5	3	3
	Cualedro	13 September 2020	25 September 2020 26 September 2020	6	3	-
	Chandrexa	13 September 2020	29 September 2020	2	3	-
	Cernado	14 September 2020	29 September 2020	3	-	2
	Cadavos	14 September 2020	30 September 2020	1	2	3
	Arbo	31 July 2022	22 August 2022	3	1	2
	Baltar	6 August 2022	16 August 2022	5	-	-
	Boiro	6 August 2022	10 August 2022 12 August 2022	16	-	5
	Carballeda de Valdeorras	22 July 2022	2 August 2022 3 August 2022	27	56	3
	Folgoso do Courel	23 July 2022	27 July 2022 28 July 2022 29 July 2022 4 August 2022	47	12	30
	Irixo	11 August 2022	29 August 2022	3	-	4
	Laza-Chandrexa	15 August 2022	18 August 2022	3	-	-
	Lobeira	26 August 2022	6 September 2022	3	-	-
	Oimbra-Rabal	21 July 2022	18 August 2022 23 August 2022	11	1	2
	Oimbra-Videferre	19 July 2022	23 August 2022	3	-	-
	Vilariño de Conso	24 July 2022	18 August 2022	2	-	-
Global	-		-	245	187	71

¹ Two forest fires occurred in the same area but in different years (July, 2020; August, 2022).

. Figure 2 shows the different soil burn severity levels classified using a modified version [6] of the soil burn severity index proposed by [43].

2.3. Radiometric Measurements

Together with the soil burn severity plots, radiometric measurements were carried out in the most homogeneous areas with the lowest percentage of tree shadows (Figure 3). The forest fires chosen for radiometric measurements were Vilaza (10 October 2018), Nerva (4 October 2018), Pedro Bernardo (2 August 2019), and Real Sitio de San Ildefonso (5 September 2019) (They were chosen by spatial proximity and good accessibility). A total of 68 radiometric measurements were taken according to the classification shown in Figure 2: 19 to low severity (Level 2), 18 to moderate severity (Level 3) and 31 to high severity (10 Level 4, 12 Level 5, and 9 Level 6).

On the one hand, the thermal radiometer CIMEL CE-312 was used on the field. This instrument is specifically designed for precise measurements of thermal infrared radiation in 6 bands (large band 8–14 µm, 8.44 µm, 8.69 µm, 9.15 µm, 10.57 µm, and 11.29 µm). CIMEL CE-312 provides real-time spectral luminance and brightness temperature measurements. The measurement procedure on the field was the following: (1) instrument installation above 1 m height tripod; (2) sky radiation correction using gold panel (perfect reflector in thermal spectral region); (3) three consecutive measurements over the same point; (4) another sky radiation correction using gold panel. Once the field measurements were performed, Land Surface Temperature (LST) and emissivity (ɛ) parameters were obtained by the temperature and emissivity separation (TES) algorithm [53]. This algorithm relies on an empirical relationship between spectral contrast and minimum emissivity, deduced from emissivity spectra obtained in both laboratory and field settings [53]. On the other hand, the optical radiometer Analytical Spectral Devices (ASD) FieldSpec^® spectrometer, with a spectral resolution range between 350 and 2500 nm and a bandwidth range between 3 and 6 nm, was used. Atmospheric absorption regions from spectrum were deleted in order to avoid noise values. Measurements were carried out in two different scenarios: the first was directly on the field; the second was at the National Institute for Aerospace Technology (INTA) laboratory. All measurements were performed on the same samples independently of the scenario. In both scenarios, three measurements at nadir per sample were taken at a distance of 20 cm between the instrument and the sample, with the sky free of clouds and avoiding tree shadows. The measurements were taken between 10:30 a.m. and 14:00 p.m. on clear, sunny day. Sky radiation correction was carried out using a Spectralon (perfect reflector in optical spectral region), calibrating and converting radiance to reflectance.

Figure 4 shows the CIMEL CE-312 emissivity data obtained after TES algorithm application. Due to the low separability among several soil burn severity levels (related to the impossibility to detect differences in soil depth), the decision was made to reclassify the measurements into three new levels: low severity (severity 2), moderate severity (3), and high severity (4, 5, and 6). Figure 5 illustrates the reclassified emissivity values.

Figure 6a shows the ASD reflectance data obtained on the field and Figure 6b at laboratory. As in the case of CIMEL CE-312 graph (Figure 4), due to the low separability among several soil burn severity levels, the same reclassification of three new levels was realized (Figure 7a,b): low severity (severity 2), moderate severity (3), and high severity (4, 5, and 6).

2.4. Radiometric Measurements Conversion to Remote Sensing Data

One of the objectives of this study was to analyze the correlation between radiometric measurements and satellite images. After analyzing the emissivity and reflectance curves obtained from the radiometric measurements (Section 2.2), the characteristics of different sensors that comply the spectral and spatial requirements were studied.

In terms of thermal data, the highest CIMEL CE-312 emissivity differences were in 10.57 and 11.29 µm (Figure 5). Close to this spectral region, several sensors such as the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) and the Moderate Resolution Imaging Spectroradiometer (MODIS), onboard TERRA/AQUA satellite, VIIRS (Visible Infrared Imaging Radiometer Suite, second generation of MODIS) onboard SNPP, NOAA-20, and NOAA-21 satellites and the Thermal Infrared Sensor (TIRS), onboard Landsat-8/9 satellite, are commonly used in multiple forest fires applications [36,37,38,39,40]. However, TIRS only has two bands in thermal infrared region whose emissivity is impossible to retrieve (at least three separate TIR bands are required) [39,53]. ASTER and MODIS are characterized to have several bands located in similar spectral region than CIMEL CE-312 (≈8–14 µm). Despite the good spectral correlation of both sensors with CIMEL CE-312, the low spatial resolution in TIR bands (ASTER = 70 m; MODIS = 1000 m) compared to the size of the field plots (20 m) hinders proper results validation. Considering these factors, we decided to decline thermal sensors in order to assess soil burn severity in this study.

Regarding optical data, the highest ASD reflectance (measured on the field and INTA facilities) differences were between NIR (780 nm to 1400 nm) and SWIR (1400 to 3000 nm) spectral regions (Figure 7a,b). RGB spectral regions are useful in order to mitigate the impact of atmospheric scattering due to aerosols [35]. The Red Edge and near-infrared (NIR) show low reflectance values related to vegetation decrease after forest fire [29,30,33]. Furthermore, shortwave infrared (SWIR) spectral regions show increased reflectance in the presence of ash and charred materials, providing insights into the severity of soil burn and the extent of fire-induced changes [32,34,54].

The MultiSpectral Instrument (MSI), on board the Sentinel-2A and B satellites, and the Operational Land Imagery (OLI) sensor (Landsat-8 and 9 satellites) provide images in the same spectrum. Many authors published scientific papers used one or both sensors for forest fires applications [28,55,56,57]. Although the OLI sensor is widely used, some limitations were found in this soil burn severity study. The spatial resolution was a little bit larger (30 m) than the size plot (20 m). In terms of temporal resolution, the revisit time is 8 days (even though Landsat-9 launch change revisit time to 8 days, all images before February 2022 have a revisit time of 16 days). Finally, the spectral resolution is restricted with only one band in NIR region. In contrast, MSI is characterized for having bands in the visible, near-infrared, and short-wave infrared ranges (

Table 2. Sentinel-2 bands with their respective central wavelengths and spatial resolutions. In bold font the spectral bands used in this study.

Sentinel-2 Bands	Central Wavelength (μm)	Resolution (m)
B1—Coastal aerosol	0.433	60
B2—Blue	0.490	10
B3—Green	0.560	10
Band 4—Red	0.665	10
B5—Vegetation Red Edge	0.705	20
B6—Vegetation Red Edge	0.740	20
B7—Vegetation Red Edge	0.783	20
B8—NIR	0.842	10
B8A—Vegetation Red Edge	0.865	20
B9—Water Vapor	0.945	60
B10—SWIR—Cirrus	1.375	60
B11—SWIR 1	1.610	20
B12—SWIR 2	2.190	20

). The temporal resolution is 5 days (3 days when tiles are overlapped) and spatial resolution is the same or higher than field plots (20 m). Considering the comparison between both sensors (spatial, temporal and spectral resolutions), MSI was chosen. All MSI bands used on this study have been resampled (from 10 to 20 m) in order to have the same extent as field plots. Cloud correction was carried out using the Scene Classification Image (SCL) mask [58].

ASD measurements were convolved to Sentinel-2 spectral bands configuration using the spectral response function of Sentinel-2 [59]. The reflectance obtained using spectral response function is equivalent to in situ measurements using the MSI sensor instead of ASD. Figure 8a (field) and Figure 8b (laboratory) display ASD reflectance after Sentinel-2 spectral response function.

2.5. Spectral Indices and Statistical Analysis

From the MSI reflectance obtained as ASD measurements and Sentinel-2 convolution (Section 2.4), some of the most widely used spectral indices in the study of forest fires were obtained: Burned Area Index for Sentinel-2 (BAIS2) [20], Mid-Infrared Burn Index (MIRBI) [31], Normalized Burn Ratio (NBR) [60], and Normalized Burn Ratio 2 (NBR2) [33]. In addition, related to soil assessment, several spectral indices were also used: Clay Ratio (CR) [61] and Iron Oxide ratio (IOR) [62]. Due to the impossibility of conducting radiometric measurements before the fire, all spectral indices were calculated using post-fire imagery as close as possible to the extinction dates. The algorithms of all the spectral indices described before are summarized in

Table 3. Spectral indices for soil burn severity assessment used in this study. Sentinel-2 bands are described in Table 2.

Spectral Index	Algorithm	Reference
Normalized Burn Ratio (NBR)	$NBR={\displaystyle \frac{(B8-B12)}{(B8+B12)}}$	[60]
Normalized Burn Ratio 2 (NBR2)	$NBR2={\displaystyle \frac{(B11-B12)}{(B11+B12)}}$	[33]
Mid-Infrared Burn Index (MIRBI)	$MIRBI=10\ast B11-9.8\ast B12$	[31]
Burned Area Index for Sentinel-2 (BAIS2)	$BAIS2=\left(1-\sqrt{{\displaystyle \frac{B6\ast B7\ast B8A}{B4}}}\right)\ast \left({\displaystyle \frac{B12-B8A}{\sqrt{B12-B8A}}}+1\right)$	[20]
Clay Ratio (CR)	$CR={\displaystyle \frac{B11}{B12}}$	[61]
Iron Oxide Ratio (IOR)	$IOR={\displaystyle \frac{B4}{B2}}$	[62]

.

Spectral indices (Table 3) were assessed using the separability index (Equation (1)) [63]. This algorithm describes the separability between two classes being one the minimum value recommended (if the value of this ratio is greater than or equal to 1, the two distributions are considered reasonably separated because the distance between the means is at least equal to the magnitude of the dispersion) [64].

$$SI={\displaystyle \frac{\left|{\mu }_1{-\mu }_2\right|}{{\sigma }_1+{\sigma }_2}}$$

(1)

where μ1 and μ2 are the mean values, and σ1 and σ2 are the standard deviation values of each soil burn severity level, respectively, measured by ASD. Numerous authors use SI as a statistical analysis method in considering the same importance for all classes analyzed [20,39]. In this study, high severity distinction is considered crucial as being closely related to future erosion risks [49]. For this reason, different weights were applied depending on the soil burn severity level (the higher level, the higher weight) [49]. Equation (2) shows the weighted separability index (SIw) with the different weights considered for each severity class.

$${SI}_w={(SI}_{L-M}\ast 0.4)+{(SI}_{M-H}\ast 0.6)$$

(2)

${SI}_{L-M}$ and ${SI}_{M-H}$ are the separability index (Equation (1)) between low and moderate severity and moderate and high severity, respectively. All the spectral indices whose value were higher than one were considered potential soil burn severity predictors. Furthermore, histogram graphs of spectral indices values were generated considering the lines match the optimal thresholds [65].

Soil burn severity maps obtained by the spectral indices optimal thresholds were validated using the field plots summarized in Table 1. Different metrics which are derived from the confusion matrix such as sensitivity, specificity, accuracy, balanced accuracy, F1-score and Cohen kappa index, were applied in this study [65]. Sensitivity (opposite to omission errors) and specificity (opposite to commission errors) represent the percentage of true positives and negatives, respectively [66]. Accuracy (ACC), one of the most popular statistic metrics used, is defined as a ratio between the correctly classified samples to the total number of samples. However, ACC is sensible to imbalanced data. In this sense and due to the impossibility of finding balanced data on field campaigns, the metrics F1-score and balanced accuracy (BACC) were also used. F1-score represents the harmonic mean of precision and sensitivity and BACC the mean of sensitivity and specificity [65]. Finally, the Cohen kappa index (k) was obtained to assess the classification concordance level [67,68]. The higher the k, the stronger the concordance [69,70]. All statistical metrics were conducted using the caret package by R [71].

3. Results

Moderate soil burn severity was always the highest reflectance in the whole region of the optical spectrum (350–2500 nm), except in the laboratory measurements, where high severity was slightly higher in part of the Short Wave Infrared region (≈2000–2400 nm). Low soil burn severity always showed the lowest reflectance values. Laboratory measurements showed higher separability between soil burn severity levels than field measurements. Focusing on the emissivity graphs, for 8.44 and 8.69 µm bands, the order was similar to reflectance measurements (highest emissivity for the moderate soil burn severity, followed by high and low soil burn severity, respectively). However, for bands 9.15, 10.57, and 11.29 µm, the opposite occurred (highest emissivity in the low soil burn severity, followed by high and moderate soil burn severity, respectively).

Regarding spectral indices results, Figure 9 displays the weighted separability index (SI_w) obtained for all spectral indices used in this study (Table 3). The green color represents SI_w for all soil burn severity levels measured by ASD on the field. The blue color represents SI_w for all soil burn severity levels measured by ASD on laboratory. NBR (1.06) and CM (1.07), using field measurements, and MIRBI (1.16), CM (1.11), IOR (1.05), and NBR2 (1.11), using INTA facilities measurements, reach a SI_w equal or higher to 1. Based on this criterion, BAIS2 was discarded as potential soil burn severity classifier (SI_w lower to 1).

Table 4. Optimal thresholds obtained by the lines match spectral indices histograms graphs which SI_w were equal or higher than 1 (Figure 9).

Scenario	Spectral Index	Optimal Thresholds
		Low	Moderate	High
Field	NBR	NBR < −0.55	NBR > −0.49	−0.49 ≥ NBR ≥ −0.55
	CM	CM < 0.74	CM > 0.85	0.85 ≥ CM ≥ 0.74
Laboratory	IOR	IOR < 1.39	IOR > 1.69	1.69 ≥ IOR ≥ 1.39
	CM	CM < 0.74	CM > 0.89	0.89 ≥ CM ≥ 0.74
	NBR2	NBR2 < −0.15	NBR2 > −0.06	−0.06 ≥ NBR2 ≥ −0.15
	MIRBI	−0.31 ≥ MIRBI ≥ −0.62	MIRBI > −0.31	MIRBI < −0.62

shows the optimal thresholds obtained by the lines match of spectral indices histograms graphs which SI_w were equal or higher than 1 (Figure 9). CM index has slightly different thresholds depending on the place where measurements were realized (field or laboratory).

Table 5. Results of sensitivity and specificity for each soil burn severity level obtained in this study (Low = L, Moderate = M, and High = H). Soil burn severity levels were validated using the field plots summarized in Table 1. In bold, the three highest results and in red, the highest result for each soil burn severity level and metric.

Scenario	Spectral Index	Sensitivity			Specificity			Mean (Sensitivity and Specificity)
		L	M	H	L	M	H	L	M	H
Field	NBR	0.55	0.83	0.24	0.91	0.48	0.97	0.73	0.66	0.61
	CM	0.46	0.82	0.34	1.00	0.54	0.84	0.73	0.68	0.59
Laboratory	IOR	0.83	0.52	0.70	0.78	0.92	0.84	0.81	0.72	0.77
	CM	0.46	0.67	0.59	1.00	0.66	0.73	0.73	0.67	0.66
	NBR2	0.46	0.71	0.56	0.98	0.64	0.77	0.72	0.68	0.67
	MIRBI	0.57	0.81	0.00	0.87	0.46	0.97	0.72	0.64	0.49

shows the results of sensitivity and specificity for each soil burn severity level obtained in this study (Low = L, Moderate = M, and High = H).

Table 6. Overall results of accuracy (ACC), balanced accuracy (BACC), F1-score (F1), and Cohen kappa index (k). In bold, the three highest results and in red, the highest result for each soil burn severity level and metric.

Scenario	Spectral Index	Accuracy (ACC)	Balanced Accuracy (BACC)	F1-Score (F1)	Cohen’s Kappa index (k)
Field	NBR	0.61	0.66	0.54	0.35
	CM	0.58	0.66	0.52	0.34
Laboratory	IOR	0.71	0.76	0.63	0.50
	CM	0.56	0.68	0.54	0.35
	NBR2	0.57	0.69	0.54	0.36
	MIRBI	0.57	0.61	0.42	0.28

displays the overall results of accuracy (ACC), balanced accuracy (BACC), F1-score (F1), and Cohen kappa index (k). Soil burn severity levels were validated using the field plots summarized in Table 1.

IOR spectral index represents the highest overall results (Table 6) for all metrics (ACC, BACC, F1, and k). Regarding sensitivity and specificity, IOR had the highest results but with some exceptions: IOR had the lowest sensitivity value in moderate level (0.52) and the lowest specificity value in low level (0.78).

Figure 10 shows an example of a soil burn severity map obtained by the optimal thresholds (Table 4) of the IOR spectral index (spectral index with highest accuracy results shown in Table 5 and Table 6) focused on a San Millao forest fire (field date: 10 September 2020).

4. Discussion

The intensity of soil burns significantly affects forest ecosystems, facilitating the spread of invasive exotic species even at low severity levels, thereby compromising the recovery of biodiversity and native vegetation. Moreover, severe burns negatively impact key ecosystem services, such as carbon storage and soil stability, particularly in forests with fire-sensitive species like Pinus pinea. Post-fire erosion exacerbates these effects, highlighting the necessity of implementing restoration strategies that include soil carbon recovery and vegetation cover stabilization to mitigate degradation and ensure the resilience of affected ecosystems.

This study presents, for the first time, the creation of soil burn severity maps using radiometric measurements convolved to remote sensing data. The results obtained were sufficiently novel that there were hardly any studies with similar objectives and using the same input data. However, results obtained were compared with other research studies that used other sources of information such as remote sensing or unnamed aerial vehicles (UAV’s).

Respecting field plots methodology, the composite burn index (CBI) is widely used in burn severity assessments designed to relate field plots (vegetation and soil) with satellite images [60,72,73,74]. However, several authors underline the need for separate vegetation and soil assessments due to the different fire behaviors on the two surfaces [7,47,49,50]. Ref. [6] provides a new soil burn severity classification based on [43] study, which assesses forest fires damage focused on soil. This quick and operational procedure considers soil metrics such as color, depth, structure, roots condition, and water repellence.

Field plots and radiometric measurements were taken in different places, with the exception of Nerva, Pedro Bernardo, and Real Sitio de San Ildefonso forest fires, where they were carried out jointly. This was a handicap due to the fact that the radiometric measurements were mostly taken in the Mediterranean climate region (82%) and the field plots were mostly taken in the Atlantic region (84%). Studies such as the one carried out by [75] and [76], in 2018 explain the importance of climate type and climate change in forest fires behavior assessment. Moreover, due to the heterogeneity of the burned area, the field plots are unbalanced [74]. For this reason, along with the accuracy (ACC) metric, balanced accuracy (BACC), F1-score (F1), and Cohen’s kappa index (k) were calculated in order to minimize this imbalance.

Focusing on radiometric measurements results, emissivity and reflectance values were extracted using CIMEL CE-312 (after TES algorithm application) and ASD, respectively. Previous studies consider emissivity as a good predictor of vegetation decrease [41] or fire severity indicator [77]. However, only this study explores the use of emissivity as potential soil burn severity indicator. Despite the considerable emissivity differences between the different soil burn severity levels using CIMEL CE-312 and observed in Figure 5, currently the low spatial resolution of remote sensing in the thermal spectrum made it impossible to validate the results. ASTER sensor has the most similar pixel size to field plots, but it is still too large (90 m of spatial resolution). Respecting reflectance values, ASD measurements (Figure 7a,b) showed that the bands with the highest differences between soil burn severity levels were those located in the near-infrared (NIR) and short wave infrared (SWIR) regions. Several previous studies related soil alterations caused by forest fires with SWIR region [27,57,77,78,79,80,81]. Generally, the measurements made at the INTA facilities showed the highest differences between soil burn severity levels compared to those made directly on the field. ASD measurements realized at the laboratory minimized factors such as shadows, wind, slope, among others that can influence results of field measurements considerably [82]. Nevertheless, both measurements were considered acceptable for subsequent application of the Sentinel-2 convolution filter (Figure 8a,b). Sentinel-2 imagery was chosen after the comparison with Landsat satellite. Analyzing spatial, spectral, and temporal resolution, Sentinel-2 data were more suitable to field plots configuration.

After Sentinel-2 convolution filter application to ASD measurements, a list of spectral indices was obtained. This list was based on popular spectral indices applied in forest fires research studies [20,31,33,60] and soil assessment [61,62]. Despite the high quantity of research studies used spectral indices which combines NIR and SWIR regions, Blue spectral band is useful in order to reduce atmospheric interferences [83,84]. Spectral indices which use algorithms including the blue spectral band are also useful in chlorophyll change studies [84,85].

The weighted separability index (SI_w) results obtained in Figure 9 was used to select those indices with good separability (SI_w ≥ 1). Following this criterion, only the Burned Area Index for Sentinel-2 (BAIS2) index was discarded. Clay Ratio (CR) was the only spectral indices which SI_w was equal or higher than one on field and at laboratory.

Based on the optimal thresholds in Table 4 obtained by the lines match of spectral indices histograms graphs, Sentinel-2 images were classified as soil severity maps (Figure 10) and validated with the field plots (Table 1). Analyzing soil burn severity levels individually, Table 5 displays sensitivity and specificity results for each spectral index and for each scenario. The Iron Oxide ratio (IOR) represents the highest results of specificity and sensitivity mean (0.81 in low, 0.72 in moderate, and 0.77 in high soil burn severity, respectively). However, sensitivity in moderate soil burn severity was the lowest (0.52 which corresponds to a 48% of omission error in this level). Sensitivity and specificity results obtained in this study are slightly lower than those obtained by [86] (0.94 in low, 0.73 in moderate, and 0.85 in high soil burn severity, respectively) using the Normalized Difference Water Index (NDWI) [87] validated in 80 field plots inside Villapadierna forest fire in León, Spain (2019, 88 Ha). In both studies, moderate soil burn severity had the lowest accuracy. IOR spectral index was also the best soil burn severity predictor in terms of overall metrics (ACC = 0.71, BACC = 0.76, F1 = 0.63 and k = 0.50, respectively) followed by NBR2 index (ACC = 0.57, BACC = 0.69, F1 = 0.54 and k = 0.36, respectively). In addition, IOR was the only one with acceptable k value (k = 0.50) based on [70] study table interpretation.

5. Conclusions

This study represents a novel contribution to soil burn severity assessment adding radiometric measurements combined with the use of remote sensing data. This combination provides a high accuracy and spatial representativeness, which significantly improves the assessment of soil burn severity. While satellite imagery is irreplaceable for large-scale analysis, radiometric measurements ensure accuracy and validation in local analysis. The integration of both approaches generates more robust, reliable, and useful results for environmental management and post-fire ecosystem restoration. The assessment conducted in this work enabled the fulfillment of the proposed objectives.

In thermal domain, despite the good spectral correlation of remote sensing (ASTER, MODIS) and radiometric measurements (CIMEL CE-312), the currently remote sensing low spatial resolution impedes its use. Future missions, such as The Copernicus Land Surface Temperature Monitoring (LSTM) with high spatial resolution in thermal domain, will allow new research lines in emissivity and soil burn severity correlation. In contrast, in the optical domain, Analytical Spectral Device (ASD) FieldSpec radiometer measurements are highly correlated with Sentinel-2 bands. Although measurements realized in the laboratory represent higher reflectance differences than field measurements, spectral indices were obtained in both scenarios. After Sentinel-2 filter convolution application to ASD measurements, the Iron Oxide ratio (IOR) spectral index provided the highest accuracy results: specificity and sensitivity mean (0.81 in low, 0.72 in moderate, and 0.77 in high soil burn severity, respectively), accuracy (ACC = 0.71), balanced accuracy (BACC = 0.76), F1-score (F1 = 0.63), and Cohen’s kappa index (k = 0.50). It is demonstrated that, in addition to near-infrared (NIR) and short wave infrared (SWIR) spectral bands, the use of blue bands reduce atmospheric interferences and improves the accuracy of soil burn severity mapping. Despite the difficulty and often the impracticability of organizing field campaigns on numerous occasions, future radiometric measurements should be conducted on healthy vegetation within the burned area (as representative as possible of pre-fire vegetation) allowing the analysis of change detection experienced by soil.

Classification of soil burn severity is crucial for the quantification of post-fire erosion. In this sense, the selection of a rapid, precise, and operational methodology for obtaining field plots proposed by [6], allowed for accurate soil burn severity validation. The soil burn severity maps obtained in this study could be highly valuable by forest agents for planning assessment and restoration tasks related to soil erosion effects. The methodology proposed in this work is based on freely available images and, due to its high operability (only requiring a post-fire Sentinel-2 image), can be easily adapted to any Geographic Information System (GIS).