Improving Fire Severity Analysis in Mediterranean Environments: A Comparative Study of eeMETRIC and SSEBop Landsat-Based Evapotranspiration Models

Citation: Quintano, C.; Fernández-Manso, A.; Fernández-Guisuraga, J.M.; Roberts, D.A. Improving Fire Severity Analysis in Mediterranean Environments: A Comparative Study of eeMETRIC and SSEBop Landsat-Based Evapotranspiration Models. Remote Sens. 2024,16, 361. https://doi.org/ 10.3390/rs16020361 Academic Editor: Jungho Im Received: 22 November 2023 Revised: 9 January 2024 Accepted: 10 January 2024 Published: 16 January 2024 Copyright: © 2024 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/). remote sensing Article Improving Fire Severity Analysis in Mediterranean Environments: A Comparative Study of eeMETRIC and SSEBop Landsat-Based Evapotranspiration Models Carmen Quintano 1,2,3,* , Alfonso Fernández-Manso 3,4, JoséManuel Fernández-Guisuraga 5 and Dar A. Roberts 3 1Electronic Technology Department, School of Industrial Engineering, University of Valladolid, 47011 Valladolid, Spain 2Sustainable Forest Management Research Institute, University of Valladolid, 34004 Palencia, Spain 3Department of Geography, University of California, Santa Barbara, CA 93106, USA; [email protected] (A.F.-M.); [email protected] (D.A.R.) 4 Agrarian Science and Engineering Department, University of León, Av. Astorga s/n, 24400 Ponferrada, Spain 5 Centro de Investigação e de Tecnologias Agroambientais e Biológicas, Universidade de Trás-os-Montes e Alto Douro, 5000-801 Vila Real, Portugal; [email protected] *Correspondence: [email protected] Abstract: Wildfires represent a significant threat to both ecosystems and human assets in Mediterranean countries, where fire occurrence is frequent and often devastating. Accurate assessments of the initial fire severity are required for management and mitigation efforts of the negative impacts of fire. Evapotranspiration (ET) is a crucial hydrological process that links vegetation health and water availability, making it a valuable indicator for understanding fire dynamics and ecosystem recovery after wildfires. This study uses the Mapping Evapotranspiration at High Resolution with Internalized Calibration (eeMETRIC) and Operational Simplified Surface Energy Balance (SSEBop) ET models based on Landsat imagery to estimate fire severity in five large forest fires that occurred in Spain and Portugal in 2022 from two perspectives: uniand bi-temporal (post/pre-fire ratio). Using-fine-spatial resolution ET is particularly relevant for heterogeneous Mediterranean landscapes with different vegetation types and water availability. ET was significantly affected by fire severity according to eeMETRIC (F > 431.35; p-value < 0.001) and SSEBop (F > 373.83; p-value < 0.001) metrics, with reductions of 61.46% and 63.92%, respectively, after the wildfire event. A Random Forest machine learning algorithm was used to predict fire severity. We achieved higher accuracy ( 0.60 < Kappa < 0.67 ) when employing both ET models (eeMETRIC and SSEBop) as predictors compared to utilizing the conventional differenced Normalized Burn Ratio (dNBR) index, which resulted in a Kappa value of 0.46. We conclude that both fine resolution ET models are valid to be used as indicators of fire severity in Mediterranean countries. This research highlights the importance of Landsat-based ET models as accurate tools to improve the initial analysis of fire severity in Mediterranean countries. Keywords: evapotranspiration; eeMETRIC; SSEBop; fire severity; Mediterranean 1. Introduction Mediterranean countries are frequently disturbed by wildfire events [ 1 – 5 ]. Wildfires have diverse impacts on various elements within forest ecosystems. They may contribute to the depletion of biomass through processes such as soil erosion [ 6 , 7 ] and water runoff [ 8 , 9 ]. Similarly, wildfires have effects on landscape dynamics by shaping forest composition and structure [ 10 , 11 ], negatively influencing ecosystem multifunctionality [ 12 ] and ecosystem biodiversity [ 13 ]. Alterations in land use in the last few decades (abandonment of traditional agriculture, increased unmanaged forest plantations, abandonment of forest as a resource, among others) combined with anthropogenic climate change (increasing heat waves and Remote Sens. 2024,16, 361. https://doi.org/10.3390/rs16020361 https://www.mdpi.com/journal/remotesensing Remote Sens. 2024,16, 361 2 of 16 droughts) is leading to an increase in the severity and extent of fire events, in addition to their frequency [14,15]. The assessment of fire impact is crucial for post-fire forest management, in particular to (i) identify priority areas with post-fire restoration needs [ 16 ], (ii) evaluate the efficacy of pre-fire fuel treatments [ 17 ], and (iii) characterize pre-fire conditions conducive to high fire affectation [ 18 ]. Fire severity is the most commonly used measure of fire’s effects on soils and vegetation [ 19 – 21 ]. The initial assessment of fire severity focusses on fire’s impacts on vegetation immediately after fire, whereas extended fire severity assessment aims to evaluate vegetative delayed mortality and recovery potential [ 22 ]. The excessive costs in time and resources that are required to field-measure fire severity in extensively burned areas make the use of remote sensing data and techniques more cost-effective [ 22 , 23 ]. In particular, fire severity is traditionally estimated from fire-induced changes in vegetation that are observed in red, near-infrared (NIR) and short-wave infrared (SWIR) spectral bands of remotely sensed multispectral data [ 24 – 28 ]. Specifically, the differenced Normalized Burn Ratio (dNBR [ 22 ]) computed from NIR and SWIR bands is a standard used to measure fire severity from remote sensing data. It is used in many studies, e.g., [ 24 , 26 , 29 – 31 ], and by operational programs such as the European Forest Fire Information System (EFFIS) (Rapid Damage Assessment module) and the Monitoring Trends in Burn Severity (MTBS) in the United States [32,33]. The dNBR index is usually categorized using thresholds to estimate the fire severity levels. It is possible, however, to use other classifiers. Parametric classifiers show lower accuracy and efficiency when compared to nonparametric supervised classifiers [ 34 ], like Random Forest (RF [ 35 ]). The RF algorithm offers a range of advantages, including computational efficiency, strong performance across several applications, and its ability to provide valuable insights into the importance of input variables, thereby enhancing accuracy in the classification process [ 36 ]. For this reason, the use of RF models in post-fire studies [ 37 ] is increasing and it has resulted in improved assessments of fire damage [21,27,38–41]. Apart from altering the remotely sensed signal using satellite sensors, wildfires also impact the variables of the energy balance equation [ 42 , 43 ]. In particular, latent heat flux or evapotranspiration (ET) is greatly influenced by modifications in vegetation, including changes in both species composition and structure resulting from fire events [ 44 , 45 ]. Different studies [ 46 – 52 ] have found a decrease in ET immediately after fire that can be observed for several years following the wildfire event. However, there are very few that use ET as an indicator of fire severity. Fernández-Manso et al. (2020) [ 53 ] and Quintano et al., (2020) [54] considered for the first time post-fire ET as a predictor of fire severity in Mediterranean countries. Previous studies on dNBR (and its derived indices) scarcely addressed a main drawback: the dNBR is considered as a proxy variable, which makes its eco-physical interpretation nonintuitive [ 53 ]. Hence, it would be highly beneficial to explore methodologies that, while retaining the simplicity of calculation, can be expressed in direct physical units, much like ET is (liters/day). ET is directly related to the physical and biogeochemical processes occurring in ecosystems [ 55 ], which make it more relevant for understanding the effects of fire on vegetation and the hydrological cycle [ 56 ]. ET can be used to directly assess fire’s impacts on specific aspects such as water availability, vegetation health, and soil quality [ 54 ]. Nowadays, ET can be estimated consistently over time and across different geographic areas [ 43 ], allowing the comparison of data across multiple regions (or wildfires). This capability proves crucial for evaluating long-term trends and making informed decisions regarding fire mitigation and post-fire restoration efforts. Many prior studies examining the relationship between fire effects and ET have relied on field-measured surface fluxes [ 47 , 48 , 57 , 58 ]. This approach was chosen because estimating ET accurately and at an appropriate spatial resolution from satellite data has been challenging [ 59 ]. Among the remote sensing-based ET models, the majority of ET estimation models rely on either complete or simplified versions of the energy balance equation, which consider the energy expended in converting liquid water within soil and plants into water vapor, and then subsequently liberated into the atmosphere [ 60 ]. Some Remote Sens. 2024,16, 361 3 of 16 models such as the Google Earth Engine (GEE) implementation of the Mapping Evapotranspiration at High Resolution with Internalized Calibration (eeMETRIC, [ 61 , 62 ]), the GEE Implementation of the Surface Energy Balance Algorithm for Land ( geeSEBAL, [63,64] ) and Atmosphere–Land Exchange Inverse/Disaggregation of the Atmosphere–Land Exchange Inverse (ALEXI/DisALEXI, [ 65 , 66 ]) estimate energy balance variables using both optical (shortwave) and thermal (longwave) data. In contrast, simplified approaches such as Operational Simplified Surface Energy Balance (SSEBop [67–69]) and the Priestley–Taylor Jet Propulsion Laboratory (PT-JPL [ 70 ]) do not estimate some parameters of the energy balance equation, or calculate them using simplified assumptions. ET serves also as a crucial variable within the ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station (ECOSTRESS) mission [ 71 , 72 ]. Pascolini-Campbell et al. (2022) [ 73 ] included it as a measure of pre-fire vegetation water stress, together with evaporative stress index (ESI) and water use efficiency (WUE). Their study showed that plant water stress was the dominant predictor of burn severity for the Southern California Mountain 2020 fires. Presently, the ECOSTRESS mission records data exclusively for the continental United States (CONUS) along with crucial biomes and agricultural regions. Validation sites are chosen from the FLUXNET network [74]. Accordingly, our study aims to estimate fire severity using two daily ET models, a complete one (eeMETRIC) and a simplified one (SSEBop), using an RF classification algorithm. METRIC is a widely recognized surface energy balance-based model [ 60 ], which justifies its selection. Fernandez-Manso et al. in 2020 and Quintano et al. in 2020 [ 53 , 54 ] already demonstrated that a post-fire METRIC ET image can be used to estimate fire severity in Mediterranean countries. In this paper, we validate a simplified ET model to estimate fire severity and compare its performance to that of the well-known METRIC model. As our study area has a non-complex topography, based on [ 75 ], our hypothesis is that both ET models should yield comparable estimations of fire severity. Additionally, this study can be thought of as a continuation of the first approaches made in this field by [ 53 , 54 ], as there is a need to test the use of ET in a set of study areas representing different bioclimatic situations, using new or different ET models for estimating fire severity. For this purpose, we chose five wildfires that impacted several vegetation types across the northwestern part of the Iberian Peninsula, and conducted the analysis from two perspectives: unitemporal (post-fire) and bi-temporal (post/pre-fire ratio). Fire severity estimation using the dNBR index was used as a benchmark. In this context, we sought to answer the following research questions: (1) How did the studied wildfires influenced ET both at the wildfire level and the reference plot level? (2) Can the selected ET Landsat-based models estimate fire severity more accurately than a standard methodology based on dNBR spectral index? (3) To what extent do uniand bi-temporal approaches, as well as the type of ET model (eeMETRIC/SSEBop), exert influences on the above two issues? 2. Material and Methods 2.1. Study Sites We selected five wildfires that occurred during the summer season of 2022 in the western region of the Mediterranean basin, specifically in northwestern Spain and Portugal (Figure 1). These fire events propagated under meteorological conditions marked by unprecedented drought periods and heat waves in the months preceding the wildfires. The affected sites span a wide range of environmental conditions, as described in detail in Table 1. In general, the terrain is not complex, which is relevant for the SSEBop model. Focusing on forest ecosystems, conifer forests dominated by maritime pine (Pinus pinaster Ait.) were found at lower elevations, while Scots pine (Pinus sylvestris L.) was the predominant conifer species at higher elevations. On the other hand, hardwood forests were mostly composed of species such as holm oak (Quercus ilex L.), Pyrenean oak (Quercus pyrenaica Willd.) and chestnut (Castanea sativa Mill.). Remote Sens. 2024,16, 361 4 of 16 Remote Sens. 2024, 16, x FOR PEER REVIEW 4 of 17 2. Material and Methods 2.1. Study Sites We selected five wildfires that occurred during the summer season of 2022 in the western region of the Mediterranean basin, specifically in northwestern Spain and Portugal (Figure 1). These fire events propagated under meteorological conditions marked by unprecedented drought periods and heat waves in the months preceding the wildfires. Figure 1. Location of the Courel (A), Valdeorras (B), Valdueza (C), Figueruela (D) and Vila Real (E) wildfires in the northwestern Iberian Peninsula (Portugal and Spain). The background image is a Landsat-8 false color composite (R = band 7; G = band 5; B = band 4). Figure 1. Location of the Courel (A), Valdeorras (B), Valdueza (C), Figueruela (D) and Vila Real (E) wildfires in the northwestern Iberian Peninsula (Portugal and Spain). The background image is a Landsat-8 false color composite (R = band 7; G = band 5; B = band 4). Remote Sens. 2024,16, 361 5 of 16 Table 1. Location and characteristics of the five wildfires considered in this study: Courel (A), Valdeorras (B), Valdueza (C), Figueruela (D) and Vila Real (E). Courel Valdeorras Figueruela Valdueza Vila Real Characteristics Location NW Spain NW Spain NW Spain NW Spain N Portugal Wildfire size (km2)136.12 127.35 11.86 15.00 76.41 Wildfire alarm date 14 July 2022 15 July 2022 15 July 2022 17 July 2022 17 July 2022 Elevation range (m) 500–1350 508–1525 700–930 950–1600 600–1100 Slope (%) range 20–150 10–130 0–151 0–152 0–100 Mean annual precipitation (mm) 1697 998 807 821 975 Mean annual temperature (◦C) 10.1 8.8 11.2 10.2 12.4 Plant communities (% of forest total area) Cs (20) Pp (30) Ps (30) Cs (20) Pp (15) Qi (10) Qi (25) Pp (20) Ps (5) Qp (40) Qi (30) Ps (20) Pp (30) Qr (5) Dataset Pre-fire ET scene date 8 July 2022 8 July 2022 8 July 2022 8 July 2022 8 July 2022 Post-fire ET scene date 9 August 2022 9 August 2022 9 August 2022 9 August 2022 9 August 2022 Post-fire SPOT6/7 image date 21 July 2022 24 July 2022 21 July 2022 28 July 2022 27 July 2022 23 July 2022 21 July 2022 Qp: Quercus pyrenaica; Qi: Quercus ilex; Pp: Pinus pinaster; Ps: Pinus sylvestris; Cs: Castanea Sativa; Qr: Quercus robur. 2.2. Datasets The five wildfires almost coincided in date and were located reasonably close together. For that reason, only two cloud-free images, one pre-fire (8 July 2022) and one post-fire (9 August 2022), of daily ET, with spatial resolutions of 30 m, were obtained for the eeMETRIC (EEFlux application [ 76 ]) and SSEBop [ 77 ] products. In addition, we downloaded USGS Landsat-8 Operational Land Imager (OLI) surface reflectance images (Level 2, Collection 2, Tier 1, trajectory/row 204/32) acquired on the same dates as the ET products to calculate the dNBR index. The official wildfire perimeters were obtained from Copernicus Emergency Management Service (EMS) maps, verified by the Center for Forest Fire Studies (Portugal) and by the Spanish Autonomous Governments (Junta de Castilla y León and Xunta de Galicia). SPOT 6 and 7 images provided by Copernicus-EMS at a spatial resolution of 1.5 m resolution were used to obtain ground reference fire severity data (Table 1). 2.3. Brief Comparative of eeMETRIC and SSEBop ET Models Both the eeMETRIC and SSEBop models have demonstrated their ability to estimate ET accurately in a variety of applications [ 78 ]. The two ET models are energy balance models; thus, they calculate ET by solving the surface energy balance equation ( Equation (1) ) on a pixel-by-pixel basis. The latent heat flux (energy consumed by the ET process, or heat energy used by water in its phase change from liquid to gas) is calculated as the difference between the net radiation to the surface and losses due to the sensible heat flux (energy used to heat the air) and ground heat flux (energy stored in the surface) [75]. LE =Rn−G−H(1) where LE represents latent heat flux (energy consumed by ET) (W m −2 ); R n is net radiation flux density (W m −2 ); Gis ground heat flux density (W m −2 ); His sensible heat flux density (W m −2 ). In both ET models R n ,Gand Hare derived from satellite data. Specifically, R n is estimated from surface land temperature (LST) and reflectance data; Gfrom LST, vegetation indices and previously calculated R n and Hfrom surface roughness, wind speed and LST ranges [ 61 ]. More detailed information about the METRIC model can be found in [ 61 , 79 , 80 ]. Similarly, detailed information about the SSEBop model can be found in [80–83]. Remote Sens. 2024,16, 361 6 of 16 Both ET models share numerous similarities, such as their theoretical foundation in the surface energy balance model and the capacity to incorporate both fine- (30 m) and coarse (1 km)-spatial-resolution data from Landsat and MODIS [ 78 ]. Nevertheless, they diverge significantly in terms of data prerequisites, implementation complexity, and associated costs. In summary, the eeMETRIC model is more accurate than the SSEBop model, but it has also a more complex mode of implementation, and thus, higher associated costs [ 75 ]. Regarding the eeMETRIC ET model, its main advantages are: (i) it computes all components of the energy balance model; (ii) it needs almost no ground-based measurements; and (iii) it is applicable to complex terrain. Its main limitations include: (i) it is very costly in terms of both computational time and monetary resources; (ii) it is owned by a University (Idaho, in particular) rather than by a state agency; and (iii) its uncertainty arises from the user’s choice of hot and cold reference pixels. In contrast, the main advantages of the SSEBop ET model include: (i) no ground-based measurements or manual selections of hot and cold reference pixels are required; (ii) minimal computational time; (iii) it is cost-effective; and (iv) it is owned by the United States Geological Survey (USGS). Among the main limitations of the SSEBop model, we can identify: (i) it is not applicable to complex terrain and (ii) it does not calculate sensible or ground heat fluxes of the surface energy balance equation. A more complete comparison of these two models can be found in [ 75 , 78 ]. 2.4. Data Analysis First, we computed the dNBR index from preand post-fire Landsat-8 OLI data using bands 5 (NIR) and 7 (SWIR). dNBR will be used as a benchmark to compare the ET performance in four scenarios, defined by the combination of two perspectives (uniand bi-temporal) and two ET models (eeMETRIC and SSEBop). Similarly, the post-/pre-fire ratio was calculated for both eeMETRIC ET and SSEBop ET. Second, a stratified random experimental design was adopted [ 84 ]. We defined 355 plots of 30 m × 30 m distributed over the five wildfires as ground reference data (111 low-fire-severity plots, 114 moderate-fire-severity plots, and 130 high-fire-severity plots). We used the coincident eeMETRIC/SSEBop/Landsat grid to randomly establish the plots. The classification of each reference plot was established through visual inspection based on the post-fire SPOT6/7 images (Table 1), relying on the predominant fire severity level observed within the plot. Specifically, a reference plot was assigned to the low-fireseverity category if its dead tree proportion was less than 50%, to the moderate-fire-severity category if the proportion ranged between 50 and 90%, and to the high-fire-severity category if the proportion was higher than 90% [ 23 ]. Next, values of dNBR, post-fire eeMETRIC ET, post-fire SSEBop ET, pre-/post-fire eeMETRIC ET ratio, and pre-/post-fire SSEBop ratio, at a 30 m spatial resolution, were extracted for each reference plot. Third, we conducted a statistical exploratory analysis of preand post-fire ET (eeMETRIC and SSEBop) and dNBR at both the wildfire and reference plot levels. We also implemented a one-way Analysis of Variance (ANOVA) to assess significant differences in ET products and dNBR between the fire severity categories. In this way, we could determine the number of fire severity categories that may be discriminated with statistical significance by each input. Consequently, we identified the inputs that could act as indicators of fire severity. Finally, we implemented an RF classification algorithm (univariate model) to evaluate the capacities of the two ET models (and the two temporal perspectives) to predict fire severity, using dNBR as a benchmark. The selection of the RF classification algorithm was based on its ability to effectively capture non-linear relationships between the dependent variable and predictors, as well as to address complex interactions among these variables [ 85 , 86 ]. To ensure prediction stability, the RF hyperparameter ntree was set to a value of 2000 [ 87 ]. The mtry RF hyperparameter was set to one because of the univariate nature of the models. The performance of the RF classification was evaluated by computing the average confusion matrix across 10-fold cross-validation resamples. We took into consideration the following accuracy parameters: overall accuracy (OA; %), Kappa index, user’s accuracy (UA; %), Remote Sens. 2024,16, 361 7 of 16 and producer’s accuracy (PA; %). We used a Z-test based on the Kappa index to compare the accuracy of the obtained fire severity estimates from each ET model and perspective. This allowed us to verify whether the differences in the Kappa index had statistical significance [ 84 ]. Note that zc = 1.96 at the 95% confidence level, and that the null hypothesis H0: ( κ 1 −κ 2) = 0 is rejected when Z > zc. Additionally, wall-to-wall fire severity prediction maps at the wildfire scale were generated from RF model objects and raster images of the corresponding predictors. All analyses were conducted in R [ 88 ] using the RandomForest [ 89 ], caret [ 90 ] and raster [91] packages. 3. Results Preand post-fire eeMETRIC ET estimates were higher than SSEBop estimates (Table 2). eeMETRIC and SSEBop ET estimates were consistent at both the wildfire and reference plot levels, as was the dNBR index. In the reference plots, ET was reduced by 61.46% and 63.92% in the post-fire situation with respect to the pre-fire scenario according to eeMETRIC and SSEBop estimates, respectively (Table 2). ET was significantly affected by fire severity according to eeMETRIC (F > 431.35; p-value < 0.001) and SSEBop ( F > 373.83 ; p-value < 0.001 ) post-fire and ratio metrics. The strongest relationships were seen between fire severity and eeMETRIC ET estimates, particularly for the post-fire metric. The dNBR index also showed significant differences between severity categories (F = 266.44; p-value < 0.001) . Post-fire ET and postto pre-fire ET ratio gradually significantly decreased (p-values < 0.05) from lowto high-fire-severity scenarios (Figure 2). Table 2. Summary of descriptive statistics for eeMETRIC and SSEBop ET estimates, as well as for the dNBR index, at both the wildfire and reference plot levels. Variable Mean Range Interquartile Range Wildfire-level eeMETRIC pre-fire ET (mm) 4.62 0.00–10.92 3.56–5.81 eeMETRIC post-fire ET (mm) 1.71 0.00–8.73 0.93–2.46 SSEBop pre-fire ET (mm) 3.45 0.00–5.91 2.96–4.05 SSEBop post-fire ET (mm) 0.89 0.00–5.22 0.00–1.55 dNBR 456.61 −594.20–1170.18 334.82–589.16 Reference plot-level eeMETRIC pre-fire ET (mm) 5.06 0.49–8.55 4.04–6.21 eeMETRIC post-fire ET (mm) 1.95 0.00–5.61 0.63–3.15 SSEBop pre-fire ET (mm) 3.88 1.43–5.49 3.39–4.44 SSEBop post-fire ET (mm) 1.40 0.00–4.66 0.02–2.28 dNBR 433.89 −439.24–1149.89 214.19–647.79 eeMETRIC post-fire ET and the postto pre-fire ET ratio featured a slightly higher accuracy (OA > 76.26% ± 2.19%; Kappa > 0.65 ± 0.02) than SSEBop metrics ( OA > 74.09% ±1.20%; Kappa > 0.61 ± 0.01) in predicting fire severity through RF classification models, though these differences had no statistical significance (Table 3). The accuracy of fire severity estimates using the dNBR index was rather low (OA = 63.95%; Kappa = 0.46), and it was significantly different from that of any ET-based estimate. Post-fire and ratio products had similar accuracy for eeMETRIC and SSEBop. For eeMETRIC products, the greatest confusion was observed between low and moderate fire severity categories, whereas the moderate fire severity category in SSEBop was misclassified more frequently into high fire severity. The latter category showed consistently high accuracy across all ET products (Table 3). Remote Sens. 2024,16, 361 8 of 16 Remote Sens. 2024, 16, x FOR PEER REVIEW 8 of 17 Figure 2. Boxplots depicting the relationships between ET estimates and dNBR index, and fire severity. We also show one-way ANOVA results. Lowercase letters denote significant differences between fire severity categories at the 0.05 level. eeMETRIC post-fire ET and the postto pre-fire ET ratio featured a slightly higher accuracy (OA > 76.26% ± 2.19%; Kappa > 0.65 ± 0.02) than SSEBop metrics (OA > 74.09% ± 1.20%; Kappa > 0.61 ± 0.01) in predicting fire severity through RF classification models, though these differences had no statistical significance (Table 3). The accuracy of fire severity estimates using the dNBR index was rather low (OA = 63.95%; Kappa = 0.46), and it was significantly different from that of any ET-based estimate. Post-fire and ratio products had similar accuracy for eeMETRIC and SSEBop. For eeMETRIC products, the greatest confusion was observed between low and moderate fire severity categories, whereas the moderate fire severity category in SSEBop was misclassified more frequently into high fire severity. The latter category showed consistently high accuracy across all ET products (Table 3). Figure 2. Boxplots depicting the relationships between ET estimates and dNBR index, and fire severity. We also show one-way ANOVA results. Lowercase letters denote significant differences between fire severity categories at the 0.05 level. Table 3. Accuracy metrics of fire severity RF classification through ET estimates and the dNBR index. Accuracy Parameters eeMeTRIC Post-Fire ET eeMeTRIC ET Ratio SSEBop Post-Fire ET SSEBop ET Ratio dNBR Index OA % 78.31 75.78 73.24 74.93 63.95 Kappa 0.67 0.64 0.60 0.62 0.46 σκ 0.0003 0.0003 0.0003 0.0003 0.0003 Margin of Error (CI) 0.04 0.04 0.04 0.04 0.04 Lower Bound 0.63 0.60 0.56 0.58 0.42 Upper Bound 0.71 0.68 0.64 0.66 0.50 Low 76.58 83.78 72.97 79.28 73.87 PA % Moderate 69.30 63.16 60.53 60.63 42.11 High 87.69 80.11 84.62 83.85 74.62 Low 80.95 76.86 77.89 79.28 67.77 UA% Moderate 67.52 62.07 61.61 67.65 47.53 High 85.71 86.67 79.14 76.76 72.93 Z-Test eeMeTRIC Post-Fire ET eeMeTRIC ET Ratio SSEBop Post-Fire ET SSEBop ET Ratio dNBR Index eeMeTRIC post-fire ET 1.46 1.67 1.93 * 8.48 eeMeTRIC ET ratio 1.46 1.55 0.56 * 7.02 SSEBop post-fire ET 1.67 1.55 0.99 * 5.42 SSEBop ET ratio 1.93 0.56 0.99 * 6.42 dNBRindex * 8.48 * 7.02 * 5.42 * 6.42 PA: producer’s accuracy; UA: user’s accuracy; OA: overall accuracy. * Statistical significance at the 95% confidence level. The spatial patterns of the classified fire severity maps are consistent with the performances of the individual ET and dNBR products. Maps derived from the ET metrics show Remote Sens. 2024,16, 361 9 of 16 lower noise from isolated pixels than dNBR maps for all wildfires (Figure 3). The high noise (i.e., granularity) in the dNBR maps is consistent with the high classification confusion between fire severity categories when using this product (Table 3). The area classified as high fire severity was larger in SSEBop than in eeMETRIC. In both ET products, the area classified as low fire severity was higher in the ratio than in the post-fire metric, especially for the Valdeorras and Figueruela wildfires (Figure 3). Remote Sens. 2024, 16, x FOR PEER REVIEW 10 of 17 Figure 3. Predicted fire severity categories through ET estimates and the dNBR index for Courel (A), Valdeorras (B), Valdueza (C), Figueruela (D) and Vila Real (E) wildfires. 4. Discussion 4.1. Question 1: How did the Studied Wildfires Influence ET? All of the studied wildfires resulted in a decrease in ET relative to the pre-fire situation immediately after the fire event, both at the wildfire level and the reference plot level. This finding aligns with prior research indicating that ET tends to decrease following a fire event (e.g., [43,47,49,52,54,92,93]) because of fire’s impacts on vegetation, microclimate, water infiltration and runoff processes [45,57,94,95]. While high-severity fires lead to the death of nearly all existing vegetation, low-severity fires result in minimal damage to aboveground vegetation. This reduction in the transpiration of leaf surface area increases bare ground cover, evaporation from soil, and surface runoff immediately after fire. High-severity fires may alter the equilibrium between evaporation and transpiration for extended periods, ranging from years to decades, as vegetation gradually regenerates following the fire [45,50,52,53,93,96,97]. In particular, our study observed an immediate post-fire decrease in daily ET that ranged between 61% and 74% depending on both the ET metric and study level (wildfire or plot), which agrees with the post-fire ET reduction reported by other studies: among others, Fernández-Manso et al., (2020) [53] reported an ET decrease of 75% after a large wildfire in Central Portugal; Mankin and Patel (2023) [52] observed an ET decrease of up to 63% in the four wildfires in South Central USA; and Ma et al., (2020) [50] found an ET reduction in the first year after fire of between 31% and 50% of pre-fire ET in the California’s Sierra Nevada (USA). Figure 3. Predicted fire severity categories through ET estimates and the dNBR index for Courel (A), Valdeorras (B), Valdueza (C), Figueruela (D) and Vila Real (E) wildfires. 4. Discussion 4.1. Question 1: How Did the Studied Wildfires Influence ET? All of the studied wildfires resulted in a decrease in ET relative to the pre-fire situation immediately after the fire event, both at the wildfire level and the reference plot level. This finding aligns with prior research indicating that ET tends to decrease following a fire event (e.g., [ 43 , 47 , 49 , 52 , 54 , 92 , 93 ]) because of fire’s impacts on vegetation, microclimate, water infiltration and runoff processes [ 45 , 57 , 94 , 95 ]. While high-severity fires lead to the death of nearly all existing vegetation, low-severity fires result in minimal damage to aboveground vegetation. This reduction in the transpiration of leaf surface area increases bare ground cover, evaporation from soil, and surface runoff immediately after fire. Highseverity fires may alter the equilibrium between evaporation and transpiration for extended periods, ranging from years to decades, as vegetation gradually regenerates following the fire [ 45 , 50 , 52 , 53 , 93 , 96 , 97 ]. In particular, our study observed an immediate post-fire decrease in daily ET that ranged between 61% and 74% depending on both the ET metric and study Remote Sens. 2024,16, 361 16 of 16 98. Miller, J.D.; Thode, A.E. Quantifying burn severity in a heterogeneous landscape with a relative version of the delta Normalized Burn Ratio (dNBR). Remote Sens. Environ. 2007,109, 66–80. [CrossRef] 99. Safford, H.D.; Miller, J.; Schmidt, D.; Roath, B.; Parsons, A. BAER soil burn severity maps do not measure fire effects to vegetation: A comment on Odion and Hanson (2006). Ecosystems 2008,11, 1–11. [CrossRef] 100. Collins, B.M.; Miller, J.D.; Thode, A.E.; Kelly, M.; van Wagtendonk, J.W.; Stephens, S.L. Interactions Among Wildland Fires in a Long-Established Sierra Nevada Natural Fire Area. Ecosystems 2009,12, 114–128. [CrossRef] 101. Soverel, N.O.; Perrakis, D.D.B.; Coops, N.C. Estimating burn severity from landsat dNBR and RdNBR indices across western Canada. Remote Sens. Environ. 2010,114, 1896–1909. [CrossRef] 102. Van Wagtendonk, J.W.; Root, R.R.; Key, C.H. Comparison of AVIRIS and landsat ETM+ detection capabilities for burn severity. Remote Sens. Environ. 2004,92, 397–408. [CrossRef] 103. Roy, D.P.; Boschetti, L.; Trigg, S.N. Remote sensing of fire severity: Assessing the performance of the normalized burn ratio. IEEE Geosci. Remote Sens. Lett. 2006,3, 112–116. [CrossRef] 104. Ha, W.; Kolb, T.E.; Springer, A.E.; Dore, S.; O’Donnell, F.C.; Martínez, R.; López, S.; Koch, G.W. Evapotranspiration comparisons between eddy covariance measurements and meteorological and remote-sensing-based models in disturbed ponderosa pine forests. Ecohydrology 2014,8, 1335. [CrossRef] 105. García-Llamas, P.; Suárez-Seoane, S.; Taboada, A.; Fernández-García, V.; Fernández-Guisuraga, J.M.; Fernández-Manso, A.; Quintano, C.; Marcos, E.; Calvo, L. Assessment of the influence of biophysical properties related to fuel conditions on fire severity using remote sensing techniques: A case study on a large fire in NW Spain. Int. J. Wildland Fire 2019,28, 512–520. [CrossRef] 106. Häusler, M.; Nunes, J.P.; Soares, P.; Sánchez, J.M.; Silva, J.M.N.; Warneke, T.; Keizer, J.J.; Pereira, J.M.C. Assessment of the indirect impact of wildfire (severity) on actual evapotranspiration in eucalyptus forest based on the surface energy balance estimated from remote-sensing techniques. Int. J. Remote Sens. 2018,39, 6499–6524. [CrossRef] 107. Collar, N.M.; Saxe, S.; Ebel, B.A.; Boden, K.S.; Rust, A.J.; Hogue, T.S. Linking fire-induced evapotranspiration shifts to streamflow magnitude and timing in the western United States. J. Hydrol. 2022,612, 128242. [CrossRef] 108. Wilder, B.A.; Kinoshita, A.M. Incorporating ECOSTRESS evapotranspiration in a paired catchment water balance analysis after the 2018 Holy Fire in California. Catena 2022,215, 106300. [CrossRef] 109. Pimentel, R.; Arheimer, B. Wildfire impact on Boreal hydrology: Empirical study of the Västmanland fire 2014 (Sweden). Hydrol. Earth Syst. Sci. Discuss. 2018, 1–26. [CrossRef] 110. Quintano, C.; Fernández-Manso, A.; Calvo, E.; Marcos, L. Valbuena. Land surface temperature as potential indicator of burn severity in forest Mediterranean ecosystems. Int. J. Appl. Earth Obs. Geoinf. 2015,36, 1–12. 111. Van der Ent, R.J.; Wang-Erlandsson, L.; Keys, P.W.; Savenije, H.H.G. Contrasting roles of interception and transpiration in the hydrological cycle—Part 2: Moisture recycling. Earth Syst. Dynam. 2014,5, 471–489. [CrossRef] 112. Cerda, A.; Robichaud, P.R. Fire Effects on Soils and Restoration Strategies; CRC Press: Boca Ratón, FL, USA, 2009. 113. Smith, H.G.; Sheridan, G.J.; Lane, P.N.J.; Nyman, P.; Haydon, S. Wildfire effects on water quality in forest catchments: A review with implications for water supply. J. Hydrol. 2011,396, 170–192. [CrossRef] 114. Morgan, P.; Keane, R.E.; Dillon, G.K.; Jain, T.B.; Hudak, A.T.; Karau, E.C.; Sikkink, P.G.; Holden, Z.A.; Strand, E.K. Challenges of assessing fire and burn severity using field measures, remote sensing and modelling. Int. J. Wildland Fire 2014,23, 1045–1060. [CrossRef] 115. Lentile, L.; Smith, A.; Hudak, A.; Morgan, P.; Bobbitt, M.; Lewis, S.; Robichaud, P. Remote sensing for prediction of 1-year post-fire ecosystem condition. Int. J. Wildland Fire 2009,18, 594–608. [CrossRef] 116. Lewis, S.A.; Hudak, A.T.; Robichaud, P.R.; Morgan, P.; Satterberg, K.L.; Strand, E.K.; Smith, A.M.S.; Zamudio, J.A.; Lentile, L.B. Indicators of burn severity at extended temporal scales: A decade of ecosystem response in mixed-conifer forests of western Montana. Int. J. Wildland Fire 2017,26, 755–771. [CrossRef] Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.