In this study, we employed the study of Barbero et al. (2019) and the classification of the French national forest inventory (NFI) to categorize France into three major regions (Fig. 1).

The Mediterranean forest and shrubland region (sylvoecoregions J and K of the NFI; IFN, 2023) encompasses the southeast portion of the country and surrounds the Mediterranean basin. This region is composed of low, dense forests that are dominated by Quercus species (Q. ilex, Q. pubescens, Q. suber) and Pinus halepensis. These species show strategies of resistance (cork for Q. suber) and tolerance (resprouting for Q. ilex and serotiny for P. halepensis) to cope with the frequent fire regime that occurs in this region. In addition, sclerophyllous non-forest vegetation, called maquis and garrigue, is widely distributed and predominantly affected by fire disturbances (Mouillot et al., 2003).

Due to the lack of a reliable and available fire contour dataset over the country, we developed our own fire polygon dataset with BAMTs (Burned Area Mapping Tools), a semi-automated method of fire contouring at high resolution (Bastarrika et al., 2014; Roteta et al., 2021). We processed atmospherically corrected and orthorectified images from the L2A product of ESA's Sentinel-2 mission between 2020 and 2022. The BAMTs method involves the calculation of three spectral indices: Normalized Differential Vegetation Index (NDVI) (Rouse et al., 1974), Normalized Burn Ratio (NBR) (Key and Benson, 1999), and NBR2 (García and Caselles, 1991). Each fire was first spatially and temporally located using NASA's Fire Information for Resource Management System (FIRMS) or national registration from the French Base de Données des Incendies de Forêt en France BDIFF official fire registration (BDIFF, 2023) to target a BAMTs processing zone. Subsequently, we defined the date of burning to determine a pre- and post-burn period, which enabled us to represent the pre-/post-differences in the three indices on an RGB color scale composite. The date of the fire is provided by the first hotspot detected by FIRMS. The pre-fire period thus runs from 1 January of the year of interest to the fire date, and the post-fire period lasts from the fire date to a user-defined final date. The final date can vary from a few weeks to several months after the fire, to guarantee a sufficient number of satellite images without cloud cover.

From the visual observation of the pre-/post-fire composite difference, the user manually draws a burned and unburned training zone further used in a random forest classifier (Belgiu and Drăguţ, 2016), based on 500 trees and unlimited maximum tree depth. This supervised classification was used to detect changes in the NDVI, NBR, and NBR2 composites. Finally, each produced fire polygon was visually evaluated and manually tuned until the desired visual accuracy was achieved. This process allowed us to remove commission errors (e.g., erroneous fire polygons being detected in agricultural lands) and to reprocess, if needed, for better capturing omitted burned areas by enlarging the training zones. We decided to capture the low-/mid-severity fires affecting the reflectance composite but still allowing unburned islands to remain classified as “unburned”. Understory fires are not captured, as they do not affect the upper canopy reflectances. BAMTs finally generates a burned area probability map, and the shapefile of the fire polygon is generated from pixels having a burn probability >50 %. A visual inspection follows this semi-supervised procedure to confirm that the detected perimeters were indeed burned areas. This key step is hardly provided by automated methods and helps to reach the international standard recommended by the CEOS (Committee for Earth Observation Satellites) Working Group on Calibration and Validation of remote sensing datasets (Franquesa et al., 2020). By focusing on forest fires larger than 30 ha, a total of 113 fire polygons were obtained over the 2020–2022 period of analysis (Fig. 2 shows three examples of a BAMTs fire polygons). We selected only fire polygons that dominantly occurred in forests and shrublands, in turn removing fires occurring on pastures and grasslands, and thus matching the French forest fire database BDIFF.

To obtain canopy height within each fire patch, we followed the methodology described in Schwartz et al. (2022, 2023) that yielded accurate results (MAE = 2.67 m when compared to in situ forest inventory measurements) over the Les Landes maritime pine forests in France. The method to map tree height at 10 m resolution in 2020 combines optical (Sentinel-2), SAR (Sentinel-1), and spaceborne lidar data (Global Ecosystem Dynamics Investigation (GEDI)) with deep learning methods (U-Net model). Sentinel-1 (S1) is a C-band Synthetic Aperture Radar (SAR) mission launched in 2014 by the European Space Agency (ESA). Here, we used the Ground Range Detected (GRD) scenes with dual-band cross-polarization (vertical–vertical plus vertical–horizontal bands at 10 m resolution) preprocessed in Google Earth Engine. We computed a single median composite image of France based on all S1 images from the leaf-on season (1 May to 1 October 2020) separated into ascending and descending orbits, thus creating a single composite image of France with four layers: VV_ascending, VH_ascending, VV_descending, and VH_descending at 10 m resolution. The Sentinel-2 (S2) mission provides multispectral images from the Earth's surface reflectance with a revisit interval of ∼5 d. Here, we used 10 bands of the L2A product (bottom of the atmosphere reflectance) resampled at 10 m when necessary (nearest neighbors method): B2 – blue, B3– green, B4 – red, B5-B6-B7 – red edge, B8 – near infrared (NIR), B8 A – ”narrow” NIR, B11-B12 – short-wave infrared (SWIR). Similarly to S1, we computed the median of all S2 images with less than 50 % of clouds for the same time period after applying QA60 cloud mask.

The Global Ecosystem Dynamics Investigation (GEDI) mission is a spaceborne infrared lidar on the International Space Station (ISS). It provides energy return waveforms (L1B product) and derived metrics such as canopy relative height (RH) (L2A product) that describe the vertical forest structure within 25 m diameter circular footprints. In this case, we used RH95 height metrics as the reference height in order to train our model. We downloaded all GEDI footprints available for France since the beginning of the GEDI mission and filtered them using the quality flag provided by NASA. U-Net is a fully convolutional network (FCN) widely used in deep learning for image segmentation tasks. Here, we adapted this model for pixel-wise height regression, with GEDI data as the reference height and S1/S2 images as predictors. With a series of linear operations (convolutions) and non-linear “activation” functions, these types of models can learn multiscale image features, such as image texture, that are then used to carry out height predictions. More specifically, U-Nets (Ronneberger et al., 2015) are divided into a contracting path (left) and an expansive path (right) which gives it its “U” shape and enables the model to extract relevant information at different spatial scales. In total, the network has 18 convolutional layers and ∼17 million trainable weights.

We trained the model following the method described in Schwartz et al. (2022) and obtained a canopy height map for 2020 in France at 10 m resolution (See Fig. 2). Considering that this map has been designed only on the data of the year 2020, we evaluate it to be reliable for the 2020–2022 period only, particularly in the intensively managed Landes region (Petucco and Andrés-Domenech, 2018), covered by the fast-growing P. pinaster species (Serrano-León et al., 2021). Schwartz's innovative high-resolution tree height map is the basis for our method of estimating forest aboveground biomass in the remainder of this study. Note that due to GEDI RH95 properties, the heights below ∼3 m cannot be separated from bare soil or crops.

The workflow in Fig. 3 describes how the aboveground biomass losses (AGB-L) were computed for each fire patch using the 10 m resolution tree height map of Schwartz et al. (2022), the fire polygons obtained with BAMTs, and the French NFI plot data. AGB-L is defined as all direct and indirect potential biomass losses due to fire. These potential losses then include all the biomass exposed to fire, leading to either the combustion of the plant material during the fire, resulting in the release of gases and aerosols, or the formation of standing and ground dead wood, which is then decomposed or harvested by forest managers. This definition refers to the concept of potential loss used in fire risk assessments (Chuvieco et al., 2023). The calculation of AGB-L at 10 m resolution within each fire patch is performed according to the following successive steps (Fig. 3).

Step 1: tree height within each fire patch. We cropped the tree height map at 10 m resolution from Schwartz et al. (2022) within the BAMTs fire polygons, keeping only the 10 m pixels with height values higher than 3 m.

The calculation of AGB-L described in Sect. 2.4 was applied on the 113 fire patches from BAMTs. The same calculation has been realized for the whole region outside the burned areas (see Sect. 2.1) to get a rough estimation of living biomass compared to burned one. We also computed AGB-L using two global burned area products and three global height or biomass products to compare with our high-resolution results (Table 1). The 113 fire polygons from BAMTs were compared, when available, with the national fire polygons estimates from the EFFIS (San-Miguel-Ayanz et al., 2012) and MCD64A1 burned area (Key and Benson, 2005; Giglio et al., 2018). Because the BAMTs fire dataset focused on forests and shrublands, a large amount of fires and total burned area occurring on pastures and grasslands were removed from the EFFIS and MCD64A1 data. While BAMTs relies exclusively on image analysis from the Sentinel-2 sensors, the EFFIS product relies on both hotspots obtained by the MODIS sensor and Sentinel-2 images crossover. MCD64A1 (500 m resolution) relies solely on the MODIS sensor.

The AGB-L from each patch from our study was compared with AGB-L calculated from global tree height maps at 30 m resolution delivered globally by Potapov et al. (2021) and processed according to the same method as presented in Sect. 2.4 for the Schwartz tree height map. We also used AGB-L directly sampled within each patch from the ESA CCI Biomass dataset at 100 m resolution for the year 2018 (Santoro and Cartus, 2021). It should be noted that this AGB product considers all the biomass compartments, including foliage and understory vegetation, whereas our AGB product considers only woody tree biomass. Finally, AGB-L obtained in this study was compared to the (burned) wood carbon pool of the global 500 m resolution fire emissions dataset provided by Van Wees et al. (2022), based on the GFED framework and with biomass cycling simulated by the CASA model (van Wees et al., 2022). All these alternative products have a coarser resolution than our burned area and AGB-L maps (Table 1). Overall, we compared our results with an ensemble of three burned areas maps times four AGB maps, thus 12 different estimates of AGB-L for each of the three forest regions.

To establish whether a difference exists between two distributions, we used the Kolmogorov–Smirnov test (KS test). This test compares the parameters of two distributions (mean and variances) to conclude whether their difference is significant. A p value below the 0.05 threshold indicates a significant difference between the two distributions. To perform this, we used the ks.test function in the “stat” set of the R program.

During the 2022 fire season, our method estimated a burned area of 42 520 ha, which is lower than the 48 330 ha estimated by EFFIS for the same selected fires >30 ha. Our estimation excluded the spring prescribed fires, leading to a total of 66 393 ha for the whole year as estimated by EFFIS. In addition to the significantly higher burned area, this fire season was characterized by an unusual distribution outside the highly fire-prone Mediterranean region. This southeastern region of France, composed of Mediterranean forest and sclerophyllous short vegetation (shrubland, maquis, and garrigues) has been the most affected by fires each year since 2006. According to BDIFF (2006–2019) and our BAMTs estimates (2020–2022), this region (blue area in Fig. 4a) has an average annual burned area of 4812 hayr-1 from 2006 to 2021 (standard deviation (SD) = 3855 ha), with high fire years observed in 2009 (8403 ha), 2016 (9800 ha), and 2017 (15 660 ha). Over the same period, this fire-prone part of the French territory accounted for 73.5 % (SD = 18.4 %) of the burned surface. Unlike most western European regions, the southern region of France did not experience an extreme fire season in 2022, with a burned area of 7971 ha, moderately above the mean but representing only 18.7 % of the total area burned in France that year. In contrast, the rest of the French territory, separated into two types of forests in Fig. 4a (temperate and Atlantic maritime pine), experienced a significantly different burned area. The burned area in the maritime pine forest reached 26 858 ha in 2022 (Fig. 4c) compared to an average of 494 hayr-1 (SD = 379 ha) over the 2006–2021 period. In 2022, the total number of fires larger than 30 ha detected was 15, compared to an average of one to five fire events per year observed during the previous years. In this region, the average burned surface per fire event reached 3357 ha in 2022, compared to 216 (SD = 229) over the 2006–2021 period. Another keystone result was the abnormally large area of temperate forests that burned in 2022, reaching a total of 7813 ha, which is 2.6 times larger than the maximum of all the previous years (2019, burned area = 3035 ha). The most important temperate forest fire events occurred in Brittany, the Loire valley, and the Jura regions (Fig. 4a). Although the burned area in this region was significantly higher than the previous years, it was composed of only a few small fires, with an average fire size of 174 ha and a maximum fire size of 1730 ha. We conclude here that the 2022 fire season differs significantly from previous years' fire seasons since 2006 in many aspects, including the total burned area, the fire sizes, and the region affected. Consequently, the impact on biomass loss over the territory cannot be derived from simple correspondence rules based on previous years' observations. Therefore, further investigation was conducted on the biomass affected.

The accuracy of AGB-L assessment mostly depends on the accuracy of the 10 m tree height map sampled within fire patches. This tree height product was previously validated in Schwartz et al. (2022) against NFI plot data over the maritime pine forest of the Landes, showing a MAE of 2.67 m. Here, we extended this validation effort to other forest regions using a larger set of NFI plot measurements collected during 2019–2021. Only the plot locations within a 5 km area around each fire patch were considered (n=451). The validation results presented in Fig. 5 show a good agreement between our 10 m resolution canopy height map and the NFI height values, with mean absolute errors of 2.38, 2.85, and 2.59 m for the Mediterranean (n=118 plots), maritime pine (n=86 plots), and temperate forests respectively (n=247 plots). Outliers in Fig. 5, where the predicted height is close to zero, are likely to be forests that were measured by a former NFI census and were then clear-cut during the observation period for which our height map is established, especially in intensively managed forests such as Les Landes (maritime pine forest). Removing these outliers (when the predicted height <4 m) results in an MAE of 2.08 m.

As a second step, we examined the relationship between tree height in 10 m burned pixels and biomass, calculated as described in 3.4 (Fig. 6). We observe that, depending on the region considered and even within one region, tree height and biomass are not correlated in the same way. In Mediterranean forests, trees are relatively smaller, but biomass increases faster with height due to a high tree density and hardwood species. Conversely, in the Atlantic maritime pine forest, trees can reach upper canopy heights of up to 25 m but with lower biomass due to high forest management limiting tree density and lower wood density for this species. Interestingly, for the old unmanaged forest of La Teste-de-Buch (tree height ∼25 m), we observe a different tree height/biomass correlation, with a higher biomass of highest trees compared to the neighboring managed forest. Temperate forests show an intermediate tree height/biomass relationship, with more biomass at a given tree height than in the maritime pine forest but less than the Mediterranean forests, with a bimodal relationship between needle leaf low wood density and hardwood forest with high wood density co-occurring in the region (Fig. 6).

When crossing our fire patches with our biomass map, our results reveal a marked increase in the amount of aboveground biomass loss (AGB-L) between the 2022 fire season (2.553 Mt) and the 2 preceding years (2020: 0.361 Mt, 2021: 0.805 Mt). Although the burned area in 2022 was significantly larger than the 2 previous years, being 10.3 and 4.5 times larger than in 2020 and 2021, the increase in biomass loss was proportionally lower, being only 7.1 and 3.2 times larger respectively. Between 2020 and 2021, the increase in biomass loss (times 2.21) was almost proportional to the burned area increase (times 2.38).

The 2022 fire season showed a major contribution of biomass loss from the maritime pine forests in Les Landes (68.2 %; 1.74 Mt) and by temperate forests (22.3 %; 0.57 Mt), associated with a lower relative contribution of 63.0 % and 18.3 % in burned area. The lower impact of fires on biomass loss in 2022 compared to the burned area affected is mainly due to the low contribution of biomass loss observed over Mediterranean forests, which represents only 9.8 % (0.25 Mt) of the total loss during the year 2022. In this latter region, the AGB-L was surprisingly 74 % lower than in 2021 and 42 % lower than in 2020. This decrease in AGB-L cannot be attributed to a reduction in burned area alone because in 2022 (7971 ha) about as much area was burned as in 2021 (7275 ha) and more than in 2020 (3141 ha). This difference in the Mediterranean biomass/burned area balance requires a more comprehensive analysis of biomass distribution within regions.

To investigate how the spatial heterogeneity of biomass within and across fire patches affects AGB-L, we conducted a study on the biomass distribution within burned areas. The distribution of the AGB-L density (tha-1) over burned areas during the period of 2020–2022 is shown in Fig. 8. In 2022, the distribution of AGB-L in Mediterranean forests has a left-hand skewed distribution, with a mode at 10 tha-1. This value corresponds to the sclerophyllous biomass density that we fixed and could not be derived from tree height data. This burned biomass distribution is actually different from the available biomass distribution (red in Fig. 8) over the whole Mediterranean region (KS test p value<0.001). Our findings indicate that the 2022 fires in the Mediterranean region primarily impacted low biomass vegetation, with 88.6 % of the burned area affecting biomass values under 100 tha-1. Notably, sclerophyllous short vegetation, which represents 26 % of the woody land cover, was found to be proportionally much more affected by fires (Fig. 8a for years 2020 and 2021). This vegetation type has a biomass of only 14 % of the median biomass of the entire region. Consequently, the higher propensity of shrublands to fire reduces the impact on biomass loss. This skewed distribution towards low biomass vegetation was even more prevalent in 2022 than during previous years.

Our study reveals that the distribution of AGB-L in maritime pine forests exposed to fire in 2022 shows a binomial distribution. The first and main peak from 20 to 80 tha-1 (35.5 % of the distribution) in Fig. 8b corresponds to the intensively managed mono-specific stands with heterogeneous ages. The area occupied by young stands with a low biomass is larger than in other regions due to shorter rotations and a larger harvested fraction. Notably, the main part of the burned distribution is actually close to the biomass distribution across the entire region. The second peak occurs over a very high biomass density (around 270 tha-1, 10.2 % of distribution) representing the old-growth, unmanaged forest that burned during the extreme fire of La Teste on the coast. The La Teste fire AGB-L outlier illustrates one of the distinctive facets of the 2022 fire season.

Furthermore, the distribution of AGB-L in burned temperate forests in 2022 shows a rather homogeneous distribution between 20 and 250 tha-1. The high biomass AGB-L range of 150–250 tha-1 corresponds to 14.3 % of the distribution. This distribution of AGB-L contrasts with the distribution over the entire region, indicating that the 2022 fires affected a relatively higher proportion of old-growth forests than were available to burn (KS test p value<0.001). This impact on carbon loss was greater than expected and corresponds to another distinctive facet of the 2022 fire season.

The refined BAMTs method (see Sect. 2.2) shows differences in the detection rate and total area of fire polygons compared to the EFFIS and MCD64A1 products (Table 2). In contrast to BAMTs, EFFIS detected only 89 % (101 fires) of the 113 fires due to a lower detection capacity of small fires (<100 ha), as only 51 of 63 fires were recorded (81 %). However, all fires >100 ha were detected. Despite its limited capability to detect small fires, EFFIS reported a much higher total burned area of 62 002 ha over the period of 2020–2022 compared to 56 053 ha in the BAMTs method. This 10 % overestimation by EFFIS was primarily due to the overestimation of large fires (>500 ha), with the total burned area of the 16 largest fires of 2020–2022 being 4649 ha (+10 %) higher. On the other hand, the MCD64A1 product presents a higher burned area in the same sample of 113 fires detected by BAMTs, despite its lower detection rate of only 39 % (44 fires), which is much lower than EFFIS. Although MCD64A1 was capable of detecting almost all large fires (>500 ha), it was unable to achieve the same accuracy on intermediate fires (100–500 ha), with a detection rate of 68 % (23 fires), and even less on small fires (<100 ha), with a detection capability of only 10 % (6 fires). Over the matching fire events, we observe that MCD64A1 tends to overestimate the burned area for fires larger than 500 ha by 16 % (Fig. 9).

Regarding our estimate of AGB-L (Sect. 2.4), we found higher values than the one obtained from global products (Fig. 10). When our biomass estimation method is applied to the tree height data from Potapov et al. (2021), we estimate that the AGB-L is 8 % lower. This difference reflects the propensity of our 10 m resolution tree height map to capture higher trees and their spatial variability within forest patches, including forest edges. Secondly, although the ESA CCI Biomass data take into account the fine elements of the trees (branches, leaves), it provides a much lower estimate of AGB-L (-32 %). This is particularly related to the underestimation of the biomass of Mediterranean ecosystems. Finally, the Van Wees et al. (2022) AGB-L seems relatively close to our estimates (-9 %) despite a much lower resolution. The major difference also appears in the Mediterranean region, for similar reasons as for ESA CCI Biomass. In conclusion, the lower-resolution products tend to estimate a higher burned area but a lower biomass being affected.

Our study represents a significant step towards the development of a national monitoring system for burned area polygons at high resolution for fires larger than 30 ha. This type of dataset is not yet available either for burned areas estimations delivered by forest services and previously used in scientific studies (Ruffault and Mouillot, 2017) or for regional/landscape applications relying on coarse-resolution global datasets (Barbero et al., 2019) or hand-drawn fire contours (Mouillot et al., 2003; Ganteaume and Barbero, 2019). To ensure the reliability of our dataset, we visually checked and adjusted each fire contour following international standards (Franquesa et al., 2020). We add a reminder here that pixel reflectances are mostly driven by the upper canopy so that NBR and other indices, used in BAMTs, have been shown to hardly capture understory fires (Roy et al., 2006; Morton et al., 2013). This “weakness” is shared by all global burned area datasets, as they use the dNBR (difference normalized burn ratio) index in a similar approach. Then, the mismatches we observed in burned areas across BA datasets is mostly a consequence of pixel resolution and algorithm construction. Recent advances using lidar may be of use in the future to better consider understory fires (East et al., 2023). We also set the seed of observation from registered fire events by French forest services (BDIFF), potentially non-exhaustive, thus completed by thermal anomalies (this latter source being the only source for 2022, before fire events are published by official statistics) following Majdalani et al. (2022). Our estimate of a forest burned area of 42 520 ha was lower than the official 66 363 ha provided by EFFIS for 2022, as we focused our study on forest and shrubland fires. This is a major limitation when comparing the registered national burned area's statistics with remote sensing data not filtered by land cover, as pointed out by Turco et al. (2019). EFFIS provides a total burned area assembling all sources of fires without filtering out fire types, information that is largely directly used by the media. To ensure consistency with BDIFF, the French fire observation system, we focused on forest or shrubland fires, voluntarily omitting pasture and agricultural fires where uncertainty is still high (Hall et al., 2021). We warn here about the use of raw data provided by remote sensing services without mentioning a proper description of vegetation types affected. When filtering out pastoral fires, covering roughly 15 000 ha and occurring between February and April when prescribed burning is allowed in the Pyrenees and central France, EFFIS detected 48 328 ha, much closer to our estimates, with a 10.6 % overestimation, a result actually consistent with Llorens et al. (2021). Mostly, EFFIS tends to produce a smooth external envelope of the burned polygons and ignores unburned patches within the fire patch. Still, EFFIS remains a reliable dataset when this limitation is known. We observed a much greater uncertainty on burned areas with global remote sensing datasets at coarse spatial resolution (500 m). Mostly, small fires <100 ha were not observed by MCD64A1, fires between 100 and 500 ha were underestimated, and fires larger than 500 ha were overestimated by 10 % with the same reason as for EFFIS, as they provide an external envelope omitting unburned internal patches. This caveat and the fire size threshold for reliability are consistent with Nogueira et al. (2016), or Katagis and Gitas (2022) and Galizia et al. (2021) in the Mediterranean area. To prevent misinformation to stakeholders in the future, we suggest here to increase efforts in developing a long-term nationwide observation system quickly collected and analyzed during emergency events as the 2022 fire season.

Regarding biomass, we estimated higher values than other automated global biomass information by 8 % to 32 %, mostly as a consequence of varying spatial resolution among datasets, where coarse resolution tends to underestimate higher biomass values as previously reported (Yu et al., 2022). Our AGB-L estimates rely on assumptions on tree allometry and tree density. We derived AGB only from tree height provided by remote sensing data, while standard allometry equations rely on both tree height and diameter at breast height (DBH). Field-based generic allometric equations are acknowledged to be locally variable (Henry et al., 2015) with some significant, potentially high (up to 97 %) uncertainty and impacts on C stock estimates (Vorster et al., 2020). Lidar-based biomass estimates, capturing the whole tree architecture, highly recently improved the quantification of carbon stocks in standing trees (Xu et al., 2021), fuel load for fire hazard (Fares et al., 2017), and CO2 emissions (Domingo et al., 2017) when compared to multispectral images. This latter method has been mostly based on foliage information as a proxy to tree structure and, thus, is only partially informative in logged areas where crown development saturates while tree height continues to increase (Jubanski et al., 2013). Our method, using tree height of the tallest tree at high resolution, then provides a step forward for a spatially continuous estimation of forest biomass in the highly managed French forests beside the plot-based forest inventory. The uncertainty reached in our method includes the tree height uncertainty from lidar (Schwartz et al., 2022) with an MAE varying between 2.38 and 2.85 m for tree height varying between 3 and 30 m, leading to a 10 % error. The additional error associated with the conversion of tree height to biomass is a major bottleneck in the current biomass density maps at the European or global level. Saito et al. (2022), for example, pointed out a 35 % difference between two reference global biomass datasets (GlobBiomass and GEOCARBON), and Avitabile and Camia (2018) report an error reaching up to 58 % to 67 % in Europe for biomass maps elaborated from multispectral remote sensing between 2000 and 2010 so that harmonized global AGB maps still report a 10 % to 26 % variation over Europe (Spawn et al., 2020). Using the recent methods of lidar and GEDI data, Duncanson et al. (2022) report a 40 % to 50 % RMSE from a generic method of European biomass estimates. They however also report local studies with lower errors, around 13 % to 25 %. Indeed, for example, Fassnacht et al. (2021) actually reach a 9.1 % to 15.7 % RMSE for biomass estimates from lidar in Mediterranean Chile, and Simonson et al. (2016) obtained a 10 % error in estimating Mediterranean woodland carbon stock increment from lidar compared to forest inventories. Our method then uses the benefits of the fine-resolution GEDI data with locally fitted allometric conversion rules to minimize errors in biomass estimates.

Our method also used the tallest tree height detected from GEDI data as a proxy for the forest structure at the 10 m pixel level, as suggested by Meyer et al. (2018). This constitutes a potential source of uncertainty in our study, as we had to assume a tree density function decreasing with tree height according to the generic “self-thinning” rule used in forestry in the absence of any tree density information. Tree crown delimitation from fine-resolution lidar should deliver new information to capture tree density and then reduce biomass estimate uncertainties in the future (Weinstein et al., 2021). With an average tree crown of 3 to 5 m diameter, we expect four to nine trees over the 10m×10m=100m2 pixels, close to the average of 500 trees per hectare observed in France. We expect the high resolution used in our study to reduce the error associated with the tree density function on such a low number of trees. Ultimately, it should be noted that we chose to ignore our biomass estimates for tree heights lower than 3 m and replaced it by an average shrub and grass biomass from the literature, as lidar data remain highly uncertain (>100 % error) for this forest class of tree height (Urbazaev et al., 2018; Durante et al., 2019). Shrub (Li et al., 2022) and grassland (Graux et al., 2020) biomass can be more related to foliage proxy detected by lidar or multispectral images and should be further explored for fire impact assessments. Biomass in the understory vegetation, yet hardly provided by remote sensing (Ferrara et al., 2023), would refine carbon stocks for fire impact or fire hazard assessments from this highly flammable and fire conductive vegetation layer. We also acknowledge that our study only covers a short period from 2020 to 2022, as we chose to analyze the fire impact close to the GEDI image acquisition. Tree height information is actually highly dynamic as a result of tree cover change due to disturbances or management plans (Hansen et al., 2013; Senf and Seidl, 2020), particularly in the highly managed Landes region, preventing the use of our database backward in time without increasing bias. Further, routinely updated biomass estimates would pave the way for recurrent and accurate impact assessment, as we could show the benefits of quantifying biomass loss at fine resolution for this 2022 fire season.

From our refined burned area and AGB-L estimates, we conclude a total biomass affected by fire of 2.553 Mt in 2022. This loss of biomass corresponds to 8.94 % of the biomass harvested each year (28.5 Mt) and further used for the energy (13.5 Mt), construction (9.5 Mt), and industrial (5.5 Mt) sectors (EFESE, 2023). More generally, fires impacted an amount of biomass corresponding to 3.86 % of the annual gross forest production (i.e., forest wood growth), which is 66 Mtyr-1. This fire affected biomass then contributed to a 17 % increase of the average natural mortality of all French forests, as reported by the national inventory.

We will note that most of the anomaly in 2022 comes from rarely observed fires in the Atlantic maritime pine and temperate forests. The high biomass lost in Les Landes is explained more by the large abnormal burned area. This large burned area had a lower-than-expected impact on AGB-L in the maritime pine forests because fires mostly affected managed forests with a 20/30 year exploitation return interval. The media actually thoroughly reported images of burnings in the old-growth forest of La Teste-de-Buch with 5 times more biomass and located behind the emblematic and touristic Pyla sand dune. This old-growth forest represented, however, only 10 % of the burned area in Les Landes and 30 % of the biomass loss.

Still yet poorly reported information is the low-burned area in the Mediterranean region in 2022 (Rodrigues et al., 2023). In addition, we could show in our results the low biomass affected (mostly shrublands) compared to all available burnable biomass in the region. Fire return intervals in Mediterranean ecosystems vary between 15–25 years for shrublands and 70–120 years for forests (Mouillot et al., 2002) so that 4 to 5 times more burned area in shrublands is expected, while we experience 6 times more burned area in shrublands than forests in 2022.

The most concerning impact is then the high biomass lost in temperate forests due to high biomass densities. Unmanaged and low-fire-prone temperate forests have been accumulating large amounts of biomass across the last decades and thus were highly impacted when the fire occurred. The burned area remains low and represents only 0.056 % of the temperate forest surface. Accounting for biomass loss appears as a substantial informative variable to characterize a fire season and its ecological impacts, rather than burned area alone, which is potentially misleading when affecting contrasted ecosystems, sylvoecoregions, and forest structures.

As a first attempt, we focused on estimating the AGB affected by fires without separating surviving trees, dead trees, and gaseous/aerosol emissions into the atmosphere. Future development of fire impact assessment should dig further into total biomass, including foliage and fine fuel such as understory and litter (Li et al., 2022). This biomass fraction is economically less valuable but highly combustible and contributes to carbon and pollutant emissions to the atmosphere. Mortality in Quercus species is actually low, and previous research could show the high mortality rate in pine forests (Garcia-Gonzalo et al., 2011). The high crown architecture of P. pinaster could induce a low combustion rate and lower carbon emissions compared to other species. Combustion efficiency, defined as the fraction of affected biomass emitted to the atmosphere as CO2, is usually assumed to not exceed 20 % for the woody compartment of forests (Mouillot et al., 2006), with even lower values of 0.5 % to 3.5 % in Mediterranean forests (Harmon et al., 2022), leaving a large amount of standing biomass for decomposition (Campbell et al., 2016) or harvesting. Post-disturbance harvesting is actually a dominant strategy in Les Landes to collect timber and biomass for low-value wood products or the paper industry. For 2022, the carbon impact may be minor, while the landscape value of the touristic Landes forest might have suffered more regarding its attractiveness and emotional aspects to society (Tribot et al., 2018). One final major, yet hardly assessed, impact might be soil carbon combustion. This neglected aspect in most ecosystems in Mediterranean Europe due to low carbon stock is a major concern in boreal forests and peatlands (Wiggins et al., 2021). The soil in Les Landes pine forest is actually slowly decomposed due to the chemical properties of the needles growing on a well-drained sandy soil, thus accumulating more carbon in the soil than in Mediterranean forests as suggested by the upper soil carbon stocks inventory of France (Martin et al., 2021). Rough estimates of forest soil carbon deeper than the upper 20 cm do not exist at the national or European level, and smoldering combustion is hardly detected by remote sensing as yet. We might suspect high smoldering combustion after the 2022 fire season in Les Landes as identified by local firefighters when the Landiras fire started again in a second large fire from soil combustion lasting more than 2 weeks, a phenomenon previously observed in Mediterranean pine forests (Xifré-Salvadó et al., 2020). This should be further investigated in terms of soil carbon loss and combustion impact on atmospheric emissions.

An additional step to further account for in the holistic evaluation of fire ecological impacts for the year 2022 would be the resilience capacities of the exposed species, as Qin et al. (2022) showed how the high fire season of 2019–2020 in Australia rapidly recovered from the efficient regenerating strategies of the affected vegetation. Mediterranean shrublands and Quercus species are efficient resprouters with high resistance (low combustion or mortality) and rapid recovery rates due to basal or tree bud regeneration, reducing the ecological impact of the 7971 ha affecting this region. P. pinaster regenerates from seeds and should recover from this disturbance, as the previous large fire in the region occurred more than 15 years before, thus allowing tree maturity and seed production. In addition, this highly managed forest was rapidly harvested after the fire to collect partially burned stems and will be rapidly planted for a fast and efficient recovery with a growth rate of 0.2–0.7 myr-1 (Lemoine, 1991). We might be more concerned about the temperate forests affected in 2022 with poor resistance strategies and low growth rates. This should be included in a fully holistic vulnerability framework to better drive communication with societies and political decisions (Forzieri et al., 2021). Post-fire recovery rates in each region could be evaluated from historical fire polygons and the current biomass map produced in this study for 2020, as performed by Berner et al. (2012) in the Euro–Siberian boreal forest.

We focused our study on the 2022 fire season, referenced as extreme for France and the overall European context, and compared this year to the 2 previous years. An extended comparison to the previous 6 years shows that 2022 was an exceptional fire year compared to preceding years, particularly outside the Mediterranean region. The extremely hot and long drought of 2022 can thus be viewed as a “natural experiment” to assess the biomass loss by new fire regimes that never or merely happened in the recent past. This conclusion that the fire season of 2022 never happened before was highly reported in the media. However, this conclusion was based on the raw EFFIS dataset, covering only the last 12 years and including pastoral fires. In the last decades, the Mediterranean region experienced way more total burned area and larger fire sizes, such as in the year 2003 when a heat wave covered most of Western Europe (Ciais et al., 2005), and even earlier in 1989, when fire prevention was less effective than nowadays (Ruffault and Mouillot, 2015). Regarding the Atlantic maritime pine forests, this region is historically less affected than the Mediterranean, but rare large fires actually occurred in the late 1940s, e.g., in 1949 when 130 000 ha were burned (Papy, 1950). We could argue that fire fighting services back in the 1940s were less efficient, but the burned area covered 5 times the burned area of 2022 with climatic conditions close to or less extreme than in 2022. For the temperate forest fires, little information has been registered for France, but the year 1976 was marked by a severe heat wave and drought and experienced large fires in northwestern France (Boulbin, 1978; Jean and Larue, 1999). If 2022 is part of the most extreme years over the recent decades, it cannot be claimed as “never observed before”. In the absence of any reference historical database for France and other European countries covering a longer period than the remote sensing era (a few decades), misleading short-term views may continue to spread out. We encourage historical data gathering, although incomplete (Pausas, 2004; Koutsias et al., 2013; Mouillot and Field, 2005), to better characterize high fire years as exceptional, unique, or novel. Still, the year 2022, with all the information available on biomass affected and precise fire contours (yet hardly available before), could be used as a reference fire year that could potentially more regularly happen in the future.

Actually, future climate projections show how warmer temperatures combined with more frequent and more extreme droughts in Europe could potentially increase the risk of fires in the future (Mouillot et al., 2002; Ruffault et al., 2020; Fargeon et al., 2020). As a result, fire indicators such as the Fire Weather Index (FWI) are expected to increase strongly in the coming decades when considering scenarios where carbon emissions are constantly increasing (RCP8.5) (Fig. 11). The high average seasonal FWI values currently observed in the Mediterranean region of France (26.75 in 1970–2005) are expected to continue to increase until reaching an extreme danger level during the whole summer period (38.7 in 2080–2097). Despite a lower-than-usual mean FWI level in this region in 2022 compared to the previous decade, the burned area was higher than normal (7971 ha compared to 4812 hayr-1 in 2006–2021) and could tend to increase with a prolonged fire season and repeated heat waves. In the maritime pine forest, the average seasonal FWI is expected to almost double, from an average of 11.23 over the period of 1970–2005 to 20.24 over the period 2070–2097. The year 2022 was marked by an average FWI value in the normal range (12.69) of the previous decade but reaching very high local extremes (95th percentile = 18.75). This high margin of the distribution of FWI, closely linked to the extremely large fires this year, is projected to become the norm in the years 2050–2080 and could even become the low margin in the period of 2080–2097. The temperate forest experienced abnormal FWI in 2022. Projections estimate that the mean seasonal FWI should more than double between the historical period (8.59 in 1970–2005) and the end of the century (18.01 in 2080–2097). The year 2022, with an average seasonal FWI value of 13.46, already gives us a glimpse of what may be the norm in the years 2050–2070. But the fire risk is expected to be even higher at the end of the century, exceeding the values obtained for this year, which was marked by many large fires in the north of the country. Impacts on fire-induced CO2 emissions might in addition shift toward more emitting fire events (Carnicer et al., 2022).