3.1 Mapping canopy height at high resolution (10 m)

To map canopy height in France at 10 m resolution, we adapted the methods developed and presented in Schwartz et al. (2022). The processing is based on a deep learning U-Net model (Ronneberger et al., 2015) adapted from Milesi (2022) that learns multi-scale features in S1 and S2 images to predict canopy height. This model is trained on a pixel-wise regression process with GEDI RH₉₅ height data. The RH₉₅ height means that 95 % of the energy returned to the sensor comes from photon reflections below this height. It is a widely used proxy for canopy height as it is less sensitive than RH₁₀₀ to atmospheric disturbances (Fayad et al., 2021; Potapov et al., 2021). Prior to the model training, France was divided into 10 000 km² areas, which we will refer to as “tiles” in the following, that we randomly split into 4408 (75 %) training, 887 (15 %) validation, and 589 (10 %) test tiles. The detailed training processes are described in Fig. 1: (1) random selection of a training tile; (2) random selection of a 2560×2560 m subset of this tile to reduce overfitting; (3) input of the corresponding 256×256-pixel image with 14 layers from S1 and S2 (See Table 1) to the U-Net model; (4) creation of a target height image from the GEDI RH₉₅ height; (5) calculation of the mean absolute error (MAE) loss between the model outputs and the GEDI height values; and (6) loss backpropagation (model weights are modified according to the value of the loss gradient with respect to them). With regard to (4), we used the 10 by 10 m pixel corresponding to the center of the GEDI footprint as a target. We also tried different approaches, such as using 20×20 m pixels to be closer to GEDI's footprint dimensions, but they resulted in a decline in model performance. Additionally, we tried filtering GEDI by landscape type, keeping only forest footprints, but this also resulted in worse validation metrics. The process described in (6) is a key element in the training of neural networks. We performed it here with the Adam optimizer and a learning rate of 0.01 that we decreased manually by a factor of 10 when the validation loss function stopped decreasing (after ∼20 h of training). The complete training process took approximately 24 h and was done with the Amazon AWS cloud platform on a GPU NVIDIA Tesla T4 (16 GB).

3.2 From height to wood volume and biomass

To produce an AGBD map at 30 m resolution (in Mg ha⁻¹), we derived power-law allometric equations from the French NFI plot data (See Table 1). We chose a 30 m resolution to correspond to the size of the NFI plots and to obtain a sufficient number of trees within 1 pixel so that AGBD has real significance. Every year, the French Geographical Institute (IGN) measures ∼6000 new plots for the French NFI (IGN, 2018). For each 30 m circular NFI plot, dendrometric measurements are made in concentric plots of 6, 9, and 15 m radius according to the diameter at breast height (DBH) and for trees having a minimum of 7.5 cm DBH. DBH and species are collected for all the trees, and height is measured for one tree per species and DBH class. The tree volumes are computed using allometric models involving DBH and height, and a WVD estimation (in m³ ha⁻¹), which corresponds to the main stem's volume of up to 7 cm diameter, is then derived for each NFI plot. In this study, we used ratios estimated from the CARBOFOR project (Loustau, 2010), a previous work specific to French forests, to convert the NFI wood volume to oven-dry AGBD: 0.59 Mg m⁻³ for coniferous and 0.89 Mg m⁻³ for broadleaved forests. These ratios combine an expansion factor that accounts for the volume of the whole tree, including branches (1.34 for coniferous and 1.61 for broadleaf) and a tree density factor (0.44 Mg m⁻³ for coniferous and 0.55 Mg m⁻³ for broadleaf) to obtain the oven-dry AGBD. After applying these ratios to the NFI plot WVD estimations, we compared them to the mean of the FORMS-H height in each NFI plot's 30 m circular area. Based on the dominant tree species given in the NFI data, we divided coniferous and broadleaved plots and fitted two power-law allometric equations (Fig. 2a, b) with a Huber regressor method that has the advantage of being less sensitive to outliers. Even though height and biomass are two different physical quantities, height–biomass power-law allometric equations have been widely used and showed satisfying results when no other variables were available to carry out biomass predictions (Enquist, 2002; Chave et al., 2005). The estimation of AGBD obtained using these power-law relationships (Fig. 2c) shows an MAE of 61.7 Mg ha⁻¹ and an R² of 0.4 when compared to the NFI AGBD. We observe a saturation for higher AGBD values >400 Mg ha⁻¹, explained by the broad range of AGBD values observed for a given height in Fig. 2a and b, especially for higher heights.

Figure 2(a, b) Comparison of the AGBD from NFI plot data to FORMS-H for broadleaved and coniferous forests. The red line corresponds to the power-law allometric equation that we fitted to the data and used to transform our tree height map into an AGBD map. (c) Comparison of the AGBD estimated from the allometric equations and the actual AGBD of the NFI plots. The red dashed line corresponds to the 1:1 line.

Download

To obtain the AGBD map of France, we first resampled our FORMS-H height map at 30 m resolution by taking the mean height within each 30 m pixel. Then we used the Copernicus DLT map (Table 1) to estimate the dominant leaf type within each 30 m pixel and applied these power-laws allometries (Fig. 2a, b) accordingly to obtain an AGBD map (Mg ha⁻¹) of France for 2020 at 30 m resolution (FORMS-B). We also produced a WVD map (FORMS-V) at 30 m resolution (m³ ha⁻¹) by converting FORMS-B with the volume-to-biomass ratios detailed above.

3.3 Product validation

To evaluate the accuracy of our 10 m resolution FORMS-H product, we compared it to four independent datasets: the 2 479 668 GEDI test footprints, the 5475 French NFI plots measured in 2020, and two ALS datasets (one large area of 2500 km² and one set of 20 smaller areas of 1 km² each) from an ongoing French lidar HD campaign that aims to cover the entire national territory within the next years (Table 1). GEDI test footprints were taken from the 589 test tiles (Fig. 1) and filtered with the Copernicus DLT map (Table 1) to remove non-forest data. We compared here the maximum FORMS-H value within the 30 m circular plot area to the dominant height variable provided in the NFI data. This height is computed from ∼ 7 representative trees and stands for the mean height of the 100 highest trees within a surface area of 1 ha. The ALS point-cloud data come from the French lidar HD measurement campaign (Table 1). The first ALS site is a large 2500 km² area in the north of Paris with flat terrain (Fig. 3). The other 20 sites cover smaller areas (1 km² each) and are distributed in different sites with more complex topography. We used ALS data point-cloud classification (See Table 1) to create a 50 cm resolution canopy height model that we resampled at 10 m by taking the 95th percentile of height and compared it to FORMS-H. Additionally, we conducted a comparative analysis with three height maps available globally or in Europe to assess the novelty of FORMS-H (Lang et al., 2023; Liu et al., 2023; Potapov et al., 2021). These maps will be referred to as “Lang”, “Liu”, and “Potapov”, respectively, in the following.

We evaluated FORMS-B by comparison to two independent forest inventory datasets and to aggregated NFI statistics at a larger scale. The NFI does not provide official plot-level biomass estimates but rather only volume and height, and our method for establishing the height–AGBD relationships shown in Fig. 2 would lead to large errors and a degree of circularity if AGBD inferred for individual plots was used for validation. Further, it would not be an independent validation since the NFI AGBD data are used to define the relationship for transforming height maps to AGBD maps. For these reasons, we used two smaller inventories for validation (Renecofor, GLORIE, described hereafter) that are independent of the NFI. The Renecofor permanent plot network (Ulrich, 1995) gathers 102 forest plots distributed over France (Fig. 3) that grow under different climatic and soil conditions. Measurements of numerous parameters, including DBH, are performed yearly by the French National Forest Office (ONF) to monitor and understand changes in forest ecosystems. These monitoring plots were installed in public forests, and their stands are managed by local foresters with the same thinning intensity as the surrounding stands. To derive a biomass estimation that we could use as reference data to evaluate our FORMS-B product, we used species-dependent DBH-based allometric equations and wood densities described in Forrester et al. (2017). We applied generic equations based on the tree leaf type (coniferous/broadleaf) for the tree species not covered by this study. Thus, by dividing by the plot surface (0.5 ha), we obtained an AGBD estimation for each forest plot of the Renecofor network. These values were compared here to the mean FORMS-B values within a 100 m diameter circle around the location of the plot. The GLORIE forest inventory (Motte et al., 2016; Zribi et al., 2019) includes measurements of tree heights, DBH, and tree density in 104 forest stands of maritime pine located in the Landes forest (Fig. 3). This area represents mainly private and intensively managed forests, representative of the Les Landes area. In the GLORIE dataset, AGBD estimations were derived from allometric equations applied to plot measurements of DBH (Shaiek et al., 2011, for DBH ≥10 cm and Baldini et al., 1989, otherwise). These values were compared to the mean FORMS-B values within a 50 m diameter circle around the location of the plot, which corresponds to the average plot dimensions. At a regional scale, we also evaluated our FORMS-B/V products over forest ecoregions (SER; see Sect. 2). France is categorized into 91 SER based on forest types and management practices. Every year, aggregated WVD (See Table 1) estimations are provided at the SER scale by the French forest inventory service (IGN) from the NFI plot data. To evaluate the capability of our model to carry out consistent estimation at this scale, we compared for each SER the average FORMS-V WVD estimations on forest pixels, determined with the Copernicus DLT map (Table 1), to these official data. Besides, in a report evaluating the French carbon footprint by ADEME and IGN in 2019, France's above- and below-ground carbon densities were assessed for groups of these SERs in 2014. These estimations, given in above- and below-ground carbon per hectare, were converted to AGBD with ratios described in Loustau (2010): the values were divided by 0.475 MgC Mg⁻¹ to estimate the oven-dry biomass and by a 1.3 root expansion factor to account for AGBD only. As for FORMS-V, we further compared these estimates to FORMS-B to assess its performance at a regional scale. Additionally, we compared FORMS-B to two other biomass products available globally or for Europe (Liu et al., 2023; Santoro and Cartus, 2023), which were converted when necessary to WVD maps with the factors detailed in Sect. 3.2. These maps will be referred to as “Liu” and “ESACCI”, respectively, in the following.

In this study, we used several error metrics, including the mean absolute error (MAE), the mean error (ME), the mean absolute percentage error (MAPE), and the coefficient of determination (R²). The MAE gives information about the overall error, the ME highlights the model's bias, and the MAPE computes the relative error percentage to compare the model's performances on different validation datasets. We applied MAPE only to heights >5 m and to AGBD >10 Mg ha⁻¹ to avoid infinite values and to evaluate our model's performances solely on forests. The R² score indicates the performance of a regression task. A score of 1 indicates that the predicted values perfectly fit the reference data. A score below 0 indicates that the model performs worse than a model predicting the average value. The detailed formula of these metrics can be found in Appendix A.

Figure 3Presentation of the study site with the French NFI plots (green), the Renecofor forest plots (blue), the GLORIE forest plots (orange), and ALS 1 and ALS 2 data (red) used for validation.

4.1 FORMS-H: France canopy height map at 10 m resolution (2020)

Our 2020 FORMS-H product for France at 10 m resolution is presented in Fig. 4a. The overall picture highlights the fragmentation and the variety of forest types in France with forest heights mainly ranging from 0 to 30 m. The details of height prediction presented in Fig. 4b show the ability of our map to retrieve precisely forest landscape units visible on Google Maps on a broad range of heights. Forest parcels are distinctly visible with precise borders.

Figure 4(a) FORMS-H: tree height map of France at 10 m resolution for the year 2020. (b) Examples at three different locations of height prediction (left) with the corresponding © Google map images from 2020, 2018, and 2019 (right). Brighter colors indicate higher heights.

We evaluated FORMS-H using four different datasets, including the GEDI test RH₉₅ (Fig. 5.1) height data, French NFI dominant height data collected in 2020 (Fig. 5.2), and two ALS datasets from the French lidar HD campaign (Fig. 5.3, 5.4). The comparison with the test GEDI footprints (See Sect. 3.1) yields an MAE of 4.48 m and R² of 0.33 (Fig. 5.1a). This initial validation step demonstrates the ability of our model to correctly map the GEDI RH₉₅ variable for a broad range of heights. Figure 5.1b boxplots highlight the high precision of our model with a low bias, especially in the 5–30 m height range (ME =4.8 m (5–10 m); ME $=-\mathrm{0.2}$ m (15–20 m); ME $=-\mathrm{3.8}$ m (25–30 m)). Higher heights are more challenging to predict, and the model tends to underestimate them (ME $=-\mathrm{6.7}$ m for the 30–35 height range). Conversely, FORMS-H indicates higher heights than the labeled GEDI footprints for many areas categorized as low heights (Fig. 5.1a). This discrepancy can likely be explained by the quality of GEDI data, where the labels could be wrong due to atmospheric conditions or geolocation errors. These geolocation errors should normally have a symmetric pattern, with as many points overestimated for lower heights as points underestimated for higher heights. However, as detailed in the figure caption, we plotted only the footprints geolocated in forest pixels of the Copernicus DLT map. Therefore the geolocation errors related to GEDI footprints located outside forests were excluded from this graph. Our comparison with the completely independent French NFI data (Fig. 5.2) does not reveal the same outlier pattern because these forest inventory measurements are more reliable and accurately geolocated. It yields a smaller MAE of 2.94 m and a higher R² of 0.69 (Fig. 5.2a) with a distribution of predicted data very close to the NFI distribution of heights (Fig. 5.2b). Similarly to the comparison vs. the GEDI test dataset, the boxplots show that higher heights above 25 m tend to be underestimated with a ME of −2.8 m for the 25–30 m range and of −4.5 m for the 30–35 m range of heights. The performances of the model are slightly better in coniferous forests (MAE =2.93 m; R²=0.74) compared to broadleaved forests (MAE =2.94 m; R²=0.65; see Appendix B). The validation with ALS 1 (Fig. 5.3) and ALS 2 (Fig. 5.4) data from the French lidar HD campaign (Table 1) confirms the conclusions obtained from the previous datasets. ALS 1 comparison (Fig. 5.3a) yields a result similar to the French NFI with an MAE of 3.54 m, R² of 0.61, and a comparable underestimation of higher heights (ME $=-\mathrm{3.36}$ m for the 25–30 height range). Lower heights tend to be overestimated (ME =6.47 m for the 0–5 m height range) with some low-height pixels that our model predicted much higher. This discrepancy can be attributed to forest borders, where FORMS-H has smooth height transitions while the ALS data capture sharp edges more effectively. The ALS 2 data come from a more diverse range of terrains and locations but still yield a good correlation with an MAE of 4.51 m and R² of 0.53. However, we observe an increased tendency to underestimate high height values (ME $=-\mathrm{6.13}$ m for the 25–30 height range) and overestimate low values (ME =4.82 m for the 0–5 m height range). Notably, the peak of height distribution at ∼17 m in the histogram of Fig. 5.4b is 1.8 times higher than in the ALS 2 distribution, indicating that the model tends to predict the mean value rather than spanning the whole distribution of heights.

Figure 5Comparison of FORMS-H with four reference datasets. The figure displays scatterplots (a) and histograms with boxplots (b) for each of the four datasets: (1) GEDI test RH₉₅ height data, (2) French NFI height from 2020 measurements, and (3–4) two ALS RH₉₅ heights from French lidar HD measurements. Only forest pixels from the DLT map (See Table 1) are shown. In each comparison, the scatterplot (a) shows a density plot of the predicted height plotted against the reference height, with brighter colors indicating a higher density of points. The dashed line represents the 1:1 line. The histograms with boxplots (b) display the differences between the predicted and reference height for each height range of 5 m. The median value is represented by a red line, while the lower and upper quartiles are represented by the left and right edges, respectively. The whiskers symbolize the 5th and 95th percentiles.

Download

We conducted a comparative analysis between FORMS-H and three other existing tree height maps sampled over the entire metropolitan French territory (See Sect. 3.3). A visual comparison of the three maps (Fig. 6a, b, c, d) highlights the improved ability of FORMS-H to reproduce spatial patterns visible on Google Maps and ALS data compared to the Lang and Potapov maps. FORMS-H performed well in flat terrain (Fig. 6a, b) as well as in areas with higher slopes (Fig. 6c, d, slope ∼30^∘). Lang's map, released at a spatial resolution of 10 m, appears to be coarser than FORMS-H, although it still captures some height patterns observed in ALS data. In contrast, the Potapov map failed to capture most of these patterns. Liu's map, based on PlanetScope data (3 m resolution) for predicting heights after training with ALS data in selected European areas (all outside France), captures the spatial heterogeneity well and follows the pattern observed in ALS data. Furthermore, individual tree crowns and very fine-scale landscape units are visible in this very high-resolution height dataset. To verify our visual comparison, we further quantitatively compared FORMS-H to ALS data resampled at 10 m resolution (Fig. 6e, f) and found that it outperformed the other models significantly for Lang and Potapov, with an MAE of 3.54 m (ALS 1) and 4.5 m (AL2) compared to an MAE of 4.97 m (ALS 1) and 5.57 m (ALS 2) for Lang, and 5.72 m (ALS 1) and 6.8 m (ALS 2) for Potapov. R² coefficients confirm the superiority of FORMS-H (ALS 1: 0.61; ALS 2: 0.53) compared to Lang (ALS 1: 0.27; ALS 2: 0.24) and Potapov (ALS 1: 0.07; ALS 2: −0.12). Interestingly, FORMS-H also performs better than the Liu map resampled at 10 m resolution for ALS 1 (MAE =4.76 m; R²=0.41) and similarly for ALS 2 (MAE =4.53 m; R²=0.50), even though this model was trained on higher-resolution images, with ALS as reference data, which is more precise but with an uneven spatial distribution and no data in France compared to our GEDI reference data. All maps tend to underestimate the higher heights measured by ALS.

Figure 6Visual (a, b, c, d) and quantitative (e, f) comparison of ALS RH₉₅ height data resampled at 10 m resolution to FORMS-H (10 m resolution) and to three other products at 3 m (Liu et al., 2023), 10 m (Lang et al., 2023), and 30 m resolution (Potapov et al., 2021). Panels (a) comes from ALS 1 data, while (b, c, d) are from ALS 2 data. Panels (a, b) are located on flat terrains. Panels (c, d) are located on steep terrain with 20 to 40^∘ slopes. Panels (e, f) show the comparison with the whole ALS 1 and ALS 2 datasets, in forest pixels filtered with the DLT map (Table 1). The black dashed line is the 1:1 line. We applied a uniform noise value in the [−0.5, 0.5] range to the three other height products to allow a better scatterplot density visualization due to the data type provided as integers without changing the performance metrics.

4.2 FORMS-B/V: France AGBD and WVD maps at 30 m resolution (2020)

Based on the power-law allometric equations fitted between NFI AGBD and our height estimates for each NFI plot, as described in Sect. 3.2, we derived FORMS-B (Fig. 7a), a 30 m resolution AGBD map of France in Mg ha⁻¹, and FORMS-V, a 30 m resolution WVD map of France in m³ ha⁻¹. The different colors on the map represent varying levels of AGBD, with brighter colors indicating higher values.

Figure 7(a) FORMS-B: AGBD map of France at 30 m resolution for 2020. (b) Examples at three different locations of biomass prediction (left) with the corresponding © Google map images from 2020, 2018, and 2019 (right). Brighter colors indicate higher AGBD.

To assess the performance of the 30 m resolution FORMS-B product, we compared it to two existing wall-to-wall biomass maps from global and European studies, sampled over France (See Sect. 3.3). Even though the three maps were obtained with different data sources and methods, they all mostly agree on biomass order of magnitudes and high-biomass locations (Fig. 8a, b, c, d). Moreover, FORMS-B, along with the Liu map, has a higher resolution, enabling more precise biomass estimations at the forest parcel level, whereas the global product ESACCI has a coarser resolution (100 m) that precludes detailed analyses. To gauge the accuracy of each map, we further quantitatively compared them to the two forest Renecofor and GLORIE inventory datasets (Fig. 8e, f, Sect. 3.3). The Renecofor forest plots mainly consist of mature forests, considered more challenging for biomass estimations, due to the presence of understory trees not evidently sampled from satellites. FORMS-B performs better on this “challenging test” dataset than the two other biomass maps, with an MAE of 59.7 Mg ha⁻¹ (vs. 63.7 Mg ha⁻¹ for Liu and 90.7 Mg ha⁻¹ for ESACCI) for biomass values reaching up to 430 Mg ha⁻¹. Similarly to the situation for FORMS-H, the FORMS-B product (R²=0.18) has a close performance to the Liu map (R²=0.17). FORMS-B is based on two height–biomass allometric equations for coniferous and broadleaved trees and cannot capture specificity due to different tree species, which partly explains the under- or overestimations of biomass for individual forest plots. The comparison analysis with Renecofor plots for the two other biomass maps leads to the same conclusion but with a higher MAE. The GLORIE forest plots are all located in the Landes forest, an intensive maritime pine plantation in the southwest of France. On this dataset, with AGBD ranging from 0 to 160 Mg ha⁻¹, FORMS-B has an MAE of 19.7 Mg ha⁻¹, better than the two other products with an MAE of 26.5 and 28.3 Mg ha⁻¹ for Liu and ESACCI, respectively. We observe a tendency to overestimate lower AGBD values, which can be explained here by the time difference between the date of inventory and the date of the maps presented here, given the high growth rates of young maritime pines in the study area (Lemoine, 1991). Overall, Fig. 8 highlights the capability of FORMS-B to estimate biomass density, compared to other existing biomass maps, across various types of forests in France. Still, errors are larger (MAPE =24.1 % for Renecofor and 39.9 % for GLORIE) than the ones obtained for the height map validation (MAPE =18.4 % for the NFI plots and 16.9 % for ALS 1 and 22.5 % for ALS 2), which showcases the difficulty of deriving an AGBD map only from a height-based product. Our height–biomass allometric equations for both coniferous and broadleaved forests could be refined with other parameters such as forest cover and detailed for specific species, which requires a high-resolution dominant species map to increase FORMS-B accuracy.

Figure 8Visual (a, b, c, d) and quantitative (e, f) comparison of FORMS-B to two other products (Liu et al., 2023; Santoro and Cartus, 2023). Panels (a, b) are located on flat terrains. Panels (c, d) are located on steep terrain with slopes from 20 to 40^∘. Panel (e) shows the comparison with the biomass from the whole dataset of Renecofor forest plots (2019), including mainly old-growth mature forests uniformly distributed over France. Panel (f) shows the comparison with the whole dataset of maritime pine forest plots (2016) from the GLORIE project in an intensively managed forest (Les Landes, southwest of France). Points circled in red represent outliers related to clear-cuts between the date of the inventory and 2020 and were removed from the calculation of error metrics. The red dashed line represents the 1:1 line.

Figure 9(a) Comparison of FORMS-V, Liu, and ESACCI average WVD with the WVD disclosed in 2020 NFI statistics for the 91 French SER. The AGBD products were converted into wood volume with ratios described in Sect. 3.3. The average WVDs were estimated for forest pixels only according to the Copernicus DLT map (Table 1). (b) Comparison of FORMS-B, Liu, and ESACCI average AGBD with 2014 AGBD estimations (ADEME and IGN, 2019) for 40 groups of SER. The reference SER data were converted from above- and below-ground carbon to oven-dry AGB as described in Sect. 3.3. The average AGBDs were estimated for forest pixels only according to the Copernicus DLT map (Table 1). The black dotted lines represent the 1:1 line.

Download

The agency responsible for the French NFI, IGN, produces yearly statistics of wood volume and sometimes biomass at different scales from the statistical aggregation of French NFI plots. Here we compared the 2020 WVD statistics and the 2014 AGBD statistics (most recent available AGBD statistics) to our FORMS products and to the Liu and ESACCI maps at the sylvoécorégion (SER) level (See Sect. 3.3). The AGBD maps were converted to WVD with the ratios defined in Sect. 3.3. Figure 9 shows these comparisons for WVD (a) and AGBD (b), where each point represents a SER or a SER group. Both for WVD and AGBD, our FORMS products are closer to the NFI aggregated values. For wood volume, FORMS-V has an MAE of 30.0 m³ ha⁻¹ and R² of 0.63, which is significantly better than other maps (Liu: MAE =40.0 m³ ha⁻¹ and R²=0.41; ESACCI: MAE =55.8 m³ ha⁻¹ and $R^{\mathrm{2}}=-\mathrm{0.18}$). For AGBD, FORMS-B (MAE =19.3 Mg ha⁻¹; R²=0.35) and Liu's map (MAE =22.4 Mg ha⁻¹; R²=0.17) have similar performances that outperform the global ESACCI (MAE =38.7 Mg ha⁻¹; $R^{\mathrm{2}}=-\mathrm{1.27}$) map. All the products underestimate the average WVD (ME for FORMS-V: −27 m³ ha⁻¹; Liu: −25 m³ ha⁻¹; ESACCI: −48 m³ ha⁻¹) and AGBD (ME for FORMS-B: −15 Mg ha⁻¹; Liu: −12 Mg ha⁻¹; ESACCI: −35 Mg ha⁻¹) of forests at the SER scale. Figure 8e showed that all the products underestimated large AGBD values in Renecofor, which could explain this underestimation at an aggregated regional scale. Additionally, we computed the WVD and AGBD averages on all the forest pixels from the DLT Copernicus map, which includes areas not regarded as forests by IGN and that could have lower AGBD and WVD values.