the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 04 Jun 2026
Mapping forest canopy height over Europe by integrating Sentinel-1, Sentinel-2, GEDI, and ICESat-2 data
Abstract. Timely, accurate, and spatial explicit information on forest structure, such as canopy height, is important to understand and respond to ongoing changes in forests and to support the mapping of habitat structure. The availability of spaceborne LiDAR data, such as those from GEDI, has stimulated the development of continental to global canopy height maps. Yet, while GEDI data are often used to train canopy height models, these data are lacking in northern areas. In this study, we mapped canopy height over Europe at 10 m resolution by combining Sentinel-1 and Sentinel-2 data and integrating training data from GEDI and ICESat-2. The integration of ICESat-2 and GEDI data mostly enhanced the model performance in the north of Europe, where GEDI data are lacking. The model reached a RMSE of 5.77 m and a MAE of 4.09 m based on an independent validation with ALS data over about 3,700 patches across Europe. The resulting canopy height map and validation dataset have been made publicly available at https://doi.org/10.5281/zenodo.13324731 and https://doi.org/10.5281/zenodo.18471620, respectively.
- Preprint
(2268 KB) - Metadata XML
-
Supplement
(424 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on essd-2026-329', Anonymous Referee #1, 07 Jul 2026
-
AC1: 'Reply on RC1', Wanda De Keersmaecker, 30 Oct 2026
The authors present a study aimed at producing a 10 m forest canopy height map for Europe by integrating Sentinel-1, Sentinel-2, GEDI and ICESat-2 observations. Overall, this is a good study that complements the growing landscape of continental and global canopy height products. Whilst the overall modelling framework is largely based on established approaches, the work provides an important contribution by extending training to northern Europe through the integration of ICESat-2 data and by delivering an openly available European canopy height dataset.
In my opinion, one of the strongest aspects of the manuscript is the validation strategy. The authors assembled an extensive independent validation dataset from heterogeneous airborne LiDAR datasets collected across multiple European countries. Such a large-scale independent validation effort is relatively uncommon and substantially increases confidence in both the reported accuracy and the resulting data product.
The manuscript is generally well written, the methodology is clearly described, and the data product will likely be valuable for a wide range of ecological and forestry applications. My comments below are intended to further strengthen the manuscript.
We would like to thank the anonymous referee for the positive and constructive feedback.
Comments:
- The independent ALS validation dataset is one of the principal strengths of this study. However, the ALS acquisitions span multiple years, while the canopy height map represents conditions in 2020. Although the manuscript acknowledges this limitation, I encourage the authors to discuss more explicitly how temporal mismatches may influence the reported validation statistics.
We agree. While the ALS dataset is valuable, there are several limitations associated with the dataset that are relevant for the validation or for potential users of this dataset. We added the following text in the discussion at lines 343-349 to explain this more clearly:
“While we believe that the ALS dataset provides valuable information on canopy structure, several limitations should be considered when using the data. Firstly, the ALS data across Europe is heterogeneous, having been collected using different instruments and on various dates. For instance, some data from Poland was collected in 2011, whereas parts of France were surveyed in 2023. In addition, the day of year varies across the dataset. This discrepancy in the ALS flight dates may cause relatively small discrepancies due to growth or phenology and larger ones due to tree loss. Secondly, also the point density of the ALS measurements differs across the dataset. Lower point densities reduce the capacity to capture the fine-scale spatial positioning of vegetation components, potentially leading to less reliable vegetation structure metrics.”
- Clarify the added value of this dataset. The manuscript compares the proposed product with several existing canopy height datasets and demonstrates modest improvements in validation statistics. While these improvements are encouraging, I suggest that the Discussion better emphasise the broader advantages of the dataset. In particular, the integration of ICESat-2 substantially improves training coverage in northern Europe, the extensive independent validation dataset increases confidence in the results, and the provision of uncertainty estimates enhances the usefulness of the product. These aspects arguably represent the primary contribution of the study and deserve greater emphasis than the relatively small differences in RMSE.
We agree and reduced the comparison with other canopy height maps so the discussion is more focused on the other aspects. We removed the part that discussed several factors that may have contributed to the balanced error distribution of the model and modified the order of the paragraphs to put the main discussion points in front.
Minor comments:
In the Abstract and the Introduction, "spatial explicit" should be corrected to "spatially explicit"
We corrected the text
Please clarify whether the reported RMSE and MAE are computed over all validation pixels or only over pixels classified as woody vegetation.
The validation with the ALS dataset was done using both woody and non-woody pixels. We added this to the caption of Figure 3.
Table 1 refers to the combined Sentinel-1/Sentinel-2 model as "S1S2-9B", whereas Figure 7 uses "S2S2-B9". This appears to be a typographical inconsistency and should be corrected.
Corrected
The manuscript occasionally uses the terms height estimate, height prediction, and canopy height map interchangeably. Using consistent terminology throughout would improve readability.
We removed the term height prediction from the manuscript to increase consistency. We kept the term ‘map’ to refer the European map and ‘estimate’ to refer to the model outputs.
The choice of adding random noise of ±1° to the latitude and longitude variables, together with the decision to saturate the sample weights above 35 m canopy height, would benefit from a brief justification.
We added a justification to text at lines 175-179 and 181-183:
“The purpose of the latitude and longitude features is to allow the model to be aware of the approximate location within Europe. Localizing information gives the model some biogeographical context and might therefore help the model to better interpret the signal. Yet, since the location is unique for each sample, there is some risk to overfit on this feature. To avoid this, we added noise to the latitude and longitude that was randomly sampled between -1 and 1 degrees.”
“Above canopy heights of 35m, the number of observations however became low, resulting in very high model weights. To avoid that these height bins with very few observations would lead to overfitting, we saturated the weight for heights higher than 35 m.“
A concise table summarising the main characteristics of this product and the existing canopy height datasets used for comparison (e.g. training data, spatial resolution, spatial extent, independent validation, and uncertainty estimates) would provide useful context for readers.
We added a table with characteristics of the dataset to the Methods section at line 250:
“In addition, we compared the validation metrics with several other publicly available datasets (Table 2).”Table 2: Overview of the publicly available datasets used in the comparison. (*) evaluated at 30m resolution (**) resampled to 10m using maximum value
Dataset
Resolution
Height metric
Training data
Model
Extent
Potapov et al. (2021)
30 m
RH95
GEDI
Bagged regression trees ensemble
Global (within GEDI coverage)
Lang et al. (2023)
10 m
RH98
GEDI
CNN
Global
Liu et al. (2023)
3 m (*)
CHM
ALS
U-NET
Europe
Tolan et al. (2024)
1 m (**)
CHM (ALS) and RH95 (GEDI)
ALS and GEDI
ViT Encoder and DPT decoder; CNN-based GEDI correction
Global
Citation: https://doi.org/10.5194/essd-2026-329-AC1 - The independent ALS validation dataset is one of the principal strengths of this study. However, the ALS acquisitions span multiple years, while the canopy height map represents conditions in 2020. Although the manuscript acknowledges this limitation, I encourage the authors to discuss more explicitly how temporal mismatches may influence the reported validation statistics.
-
AC1: 'Reply on RC1', Wanda De Keersmaecker, 30 Oct 2026
-
RC2: 'Comment on essd-2026-329', Anonymous Referee #2, 21 Aug 2026
The manuscript presents a new 10 m canopy height dataset for Europe by combining GEDI and ICESat-2 observations with Sentinel-1 and Sentinel-2 data. Considerable effort has been devoted to collecting and processing the training data as well as compiling an extensive independent ALS validation dataset across Europe. The resulting dataset is likely to be valuable for forest monitoring and ecosystem studies. However, I have several concerns regarding the methodological description, validation strategy, and interpretation of the results that should be addressed before publication.
Major comments
1. The manuscript states that the objective is to map canopy height of woody vegetation (trees and shrubs), while the title and abstract mainly refer to forests. It would be helpful to clarify this distinction throughout the manuscript, as the mapped product appears to include woody vegetation beyond forests. In addition, the training samples are selected based on the intersection of multiple land-cover products (e.g., ESA WorldCover 2020/2021 and the Copernicus High Resolution Forest Type layer from 2018). These products differ in acquisition year and classification methodology, which may introduce temporal inconsistencies and classification uncertainties.
2. The description of the GEDI and ICESat-2 training samples is insufficient. Please report the final number of training, validation, and testing samples for each dataset, together with their spatial distribution. A map similar to Figure 1 showing the spatial density or number of GEDI and ICESat-2 samples would help readers better understand the representativeness of the training data.
3. The manuscript mainly reports the independent validation results using the ALS dataset. Since the canopy height model was trained using GEDI and ICESat-2 observations, the evaluation on the held-out GEDI/ICESat-2 test dataset should also be presented more explicitly.
4. The independent validation is based on ALS RH98, whereas the compared canopy height products are not necessarily based on the same canopy height metric (e.g., RH95 in Potapov et al.). It would be helpful to summarize these differences in the Methods or Results to facilitate a fair interpretation of the comparisons.
5. The manuscript frequently refers to the canopy height maps from Pauls et al. in the Introduction. However, their height maps are not included in the quantitative comparison.
6. Figure 4 only presents the comparison using all validation samples. Since the manuscript emphasizes the improved performance in northern Europe, it would be helpful to include equivalent comparisons for northern and southern Europe for all reference products. In addition, please clarify whether exactly the same validation samples and filtering criteria (land cover) were applied to all products to ensure a fair comparison.
7. More methodological details are needed regarding the combined training using GEDI and ICESat-2 observations. Since the two datasets differ in footprint size, sampling characteristics, and measurement uncertainty, please clarify how the training samples from the two sensors were combined and balanced during model training. In addition, the key CatBoost hyperparameters should be reported to improve reproducibility.
8. According to Figure 3, the model trained solely with ICESat-2 achieves a very similar accuracy to the model trained with both GEDI and ICESat-2. Therefore, the additional contribution of GEDI is not entirely clear.
9. Figure 5 shows a clear saturation effect, with increasing underestimation for canopy heights above approximately 25 m. In this height range, the proposed model performs worse than Lang et al. (2023).
10. Figure 6 appears to exhibit noticeable block-like spatial patterns across large parts of Europe. It is unclear whether these patterns originate from the visualization (e.g., map resampling) or from the underlying prediction product. In addition, it would be useful to compare the canopy height distributions of different maps, which may help readers better understand where the differences originate.
11. It is not clear which dataset was used to generate the results shown in Figure 7. The reported accuracy differs from the results presented in Figures 3 and 4.
12. Overall, I believe the main strength of this dataset lies in the improved canopy height mapping over northern Europe through the integration of ICESat-2 observations. However, additional clarification of the methodology, more comprehensive validation, and improved comparisons with existing products are needed to fully demonstrate the advantages of the proposed dataset. Providing sufficient methodological details would also substantially improve the reproducibility and long-term value of this dataset.Citation: https://doi.org/10.5194/essd-2026-329-RC2 -
AC2: 'Reply on RC2', Wanda De Keersmaecker, 30 Oct 2026
The manuscript presents a new 10 m canopy height dataset for Europe by combining GEDI and ICESat-2 observations with Sentinel-1 and Sentinel-2 data. Considerable effort has been devoted to collecting and processing the training data as well as compiling an extensive independent ALS validation dataset across Europe. The resulting dataset is likely to be valuable for forest monitoring and ecosystem studies. However, I have several concerns regarding the methodological description, validation strategy, and interpretation of the results that should be addressed before publication.
We would like to thank the reviewer for the summary and constructive feedback.
Major comments
1. The manuscript states that the objective is to map canopy height of woody vegetation (trees and shrubs), while the title and abstract mainly refer to forests. It would be helpful to clarify this distinction throughout the manuscript, as the mapped product appears to include woody vegetation beyond forests. In addition, the training samples are selected based on the intersection of multiple land-cover products (e.g., ESA WorldCover 2020/2021 and the Copernicus High Resolution Forest Type layer from 2018). These products differ in acquisition year and classification methodology, which may introduce temporal inconsistencies and classification uncertainties
We agree and modified the terminology in the title and abstract. Within the data preparation phase, we mainly wanted to avoid modeling the height of tall non-woody vegetation (e.g. tall crops) or other objects (e.g. buildings).
2. The description of the GEDI and ICESat-2 training samples is insufficient. Please report the final number of training, validation, and testing samples for each dataset, together with their spatial distribution. A map similar to Figure 1 showing the spatial density or number of GEDI and ICESat-2 samples would help readers better understand the representativeness of the training data.
We added the information to the supplement S.5:
“ …. The distribution of the GEDI and ICESat-2 samples is shown in figure S2. In total, 362508, 50614, and 39807 samples were available from GEDI and 870309, 117669, and 101387 samples from ICESat-2 for training, validation, and testing, respectively. 491717, 67308, and 55869 samples in the training, validation and test dataset were sampled outside woody areas. The sampling grid had an upper left corner with coordinate 2600 000 (x), 6600 000 (y) in EPSG 3035.”
3. The manuscript mainly reports the independent validation results using the ALS dataset. Since the canopy height model was trained using GEDI and ICESat-2 observations, the evaluation on the held-out GEDI/ICESat-2 test dataset should also be presented more explicitly.
We added the validation of the canopy height map over the test dataset to the supplementary information S.6:
“The validation of the canopy height estimates from the canopy height model and other datasets available over Europe is shown in Figure S3. The results were obtained using observations inside and outside woody areas.”
4. The independent validation is based on ALS RH98, whereas the compared canopy height products are not necessarily based on the same canopy height metric (e.g., RH95 in Potapov et al.). It would be helpful to summarize these differences in the Methods or Results to facilitate a fair interpretation of the comparisons.
We added a table with characteristics of the dataset to the Methods section at line 250:
“In addition, we compared the validation metrics with several other publicly available datasets (Table 2).”
Table 2: Overview of the publicly available datasets used in the comparison. (*) evaluated at 30m resolution (**) resampled to 10m using maximum value
Dataset
Resolution
Height metric
Training data
Model
Extent
Potapov et al. (2021)
30 m
RH95
GEDI
Bagged regression trees ensemble
Global (within GEDI coverage)
Lang et al. (2023)
10 m
RH98
GEDI
CNN
Global
Liu et al. (2023)
3 m (*)
CHM
ALS
U-NET
Europe
Tolan et al. (2024)
1 m (**)
CHM (ALS) and RH95 (GEDI)
ALS and GEDI
ViT Encoder and DPT decoder; CNN-based GEDI correction
Global
5. The manuscript frequently refers to the canopy height maps from Pauls et al. in the Introduction. However, their height maps are not included in the quantitative comparison.The canopy height map of Pauls et al. is indeed not included in the comparison. While the maps are available via a viewer, we couldn’t find a repository that allowed us to download and hence include the dataset. The ECHOSAT dataset from Pauls et al. is available but was published at the moment we finalized the manuscript.
6. Figure 4 only presents the comparison using all validation samples. Since the manuscript emphasizes the improved performance in northern Europe, it would be helpful to include equivalent comparisons for northern and southern Europe for all reference products. In addition, please clarify whether exactly the same validation samples and filtering criteria (land cover) were applied to all products to ensure a fair comparison.
We agree and added the validation for all reference products for northern and southern Europe to the supplementary information S.7:
“The validation results over the ALS dataset using data north of 51.6°N (Figure S4) and south of 51.6°N (Figure S5) show that the validation metrics are worse in the north for Lang et al. (2023) and Tolan et al. (2024), while Liu et al. (2023) shows mixed results. The change in the validation results in the north and south is however smaller as compared to the GEDI-based model. No comparison can be made for Potapov et al. (2021) given the low number of observations available in the northern area.”
We used the same filtering criteria for the different datasets.
7. More methodological details are needed regarding the combined training using GEDI and ICESat-2 observations. Since the two datasets differ in footprint size, sampling characteristics, and measurement uncertainty, please clarify how the training samples from the two sensors were combined and balanced during model training. In addition, the key CatBoost hyperparameters should be reported to improve reproducibility.
We added the information about the hyperparameters and the distribution of the GEDI and ICESat-2 samples in the supplementary information S5.
“The CatBoost model (CatBoostRegressor model; CatBoost version 1.2.5) was calibrated using a learning rate of 0.1, loss_function equal to 'MultiQuantile:alpha=0.05,0.5,0.95', iterations equal to 30000, 20 early_stopping_rounds, and a random_state equal to 42. The distribution of the GEDI and ICESat-2 samples is shown in figure S2. In total, 362508, 50614, and 39807 samples were available from GEDI and 870309, 117669, and 101387 samples from ICESat-2 for training, validation, and testing, respectively. 491717, 67308, and 55869 samples in the training, validation and test dataset were sampled outside woody areas. The sampling grid had an upper left corner with coordinate 2600 000 (x), 6600 000 (y) in EPSG 3035.”
8. According to Figure 3, the model trained solely with ICESat-2 achieves a very similar accuracy to the model trained with both GEDI and ICESat-2. Therefore, the additional contribution of GEDI is not entirely clear.
The contribution of GEDI to the model is indeed rather small. This could perhaps partially be explained by the sampling strategy. Over sampling grid cells where both GEDI and ICESat-2 data were available, we used a combination of both datasets. This was done to avoid a hard split in the source of the training samples between northern and southern Europe. Yet, since we could only sample from the ICESat-2 data in northern Europe, the total contribution of the GEDI samples is smaller than the ICESat-2 samples.
9. Figure 5 shows a clear saturation effect, with increasing underestimation for canopy heights above approximately 25 m. In this height range, the proposed model performs worse than Lang et al. (2023).
Yes, this is correct and was stated in the description of the results at lines 282-285:
“In general, our models show a lower error in the low canopy height ranges (0 – 10 m) than the maps of Lang et al. (2023) and Liu et al. (2023), which tend to overestimate canopy height. Within the higher canopy height ranges (30 – 40 m), all maps tend to underestimate canopy height. The underestimation of our maps is higher than those of Lang et al. (2023), but slightly lower than those of Tolan et al. (2024) and Liu et al. (2023).”
10. Figure 6 appears to exhibit noticeable block-like spatial patterns across large parts of Europe. It is unclear whether these patterns originate from the visualization (e.g., map resampling) or from the underlying prediction product. In addition, it would be useful to compare the canopy height distributions of different maps, which may help readers better understand where the differences originate.
The block pattern is indeed a visualization effect. The comparison of the distribution of the different height maps is shown in Figure 5, showing the errors over different height bins.
11. It is not clear which dataset was used to generate the results shown in Figure 7. The reported accuracy differs from the results presented in Figures 3 and 4.
Figures 3 and 4 were based on the ALS dataset, while figure 7 is derived from the spaceborne LiDAR dataset. We added this to the caption of Figure 7 to avoid confusion.
12. Overall, I believe the main strength of this dataset lies in the improved canopy height mapping over northern Europe through the integration of ICESat-2 observations. However, additional clarification of the methodology, more comprehensive validation, and improved comparisons with existing products are needed to fully demonstrate the advantages of the proposed dataset. Providing sufficient methodological details would also substantially improve the reproducibility and long-term value of this dataset.
Citation: https://doi.org/10.5194/essd-2026-329-AC2
-
AC2: 'Reply on RC2', Wanda De Keersmaecker, 30 Oct 2026
-
RC3: 'Comment on essd-2026-329', Anonymous Referee #3, 23 Sep 2026
This manuscript presents a 10 m canopy-height product for Europe derived from Sentinel-1 and Sentinel-2 data and trained using GEDI and ICESat-2 observations. The modelling framework itself is largely based on established approaches. However, I do not consider the limited methodological novelty to be a fundamental problem for ESSD. The main community value lies in the Europe-wide data product and, particularly, in the effort devoted to compiling and processing a geographically extensive independent ALS validation dataset. The reported performance is competitive, and the annual temporal compositing and filtering appear to produce a useful product.
I therefore consider the datasets potentially suitable for publication in ESSD. However, the present version does not yet meet the journal’s expectations for data completeness, reproducibility, uncertainty characterization, and transparent quality assessment. I recommend major revision.
Review of “Mapping forest canopy height over Europe by integrating Sentinel-1, Sentinel-2, GEDI, and ICESat-2 data”
Major comments
1. Consistency of the released ALS dataset
The public validation dataset does not appear to match the dataset described in the manuscript. Supplementary Table S2 reports 3,695 patches across 18 countries, including 385 patches from Poland, whereas the Zenodo record contains 3,310 patches across 17 countries and does not include Poland. The difference is exactly 385 patches. Please reconcile this discrepancy and systematically verify that the manuscript, repository description, spatial index, and released files are mutually consistent.2. Reproducibility of the modelling workflow
The present description is insufficient to reproduce the model-derived product. Although the quantiles and general weighting approach are described, essential information such as the CatBoost version, complete hyperparameters, stopping procedure, random seeds, exact weighting implementation, and GEDI–ICESat-2 sample balance is missing. The final numbers of training, validation, and test samples for each sensor should also be reported.At minimum, the manuscript and repository should provide all parameters, software versions, spatial split information, and processing details required to reproduce the results. Archiving version-controlled code and an environment file would be the clearest way to achieve this. CatBoost should also be properly cited.
3. ALS validation design and geographic representativeness
The ALS dataset is extensive but strongly concentrated in a few countries. Spain, Norway, Finland, and Sweden account for approximately 57% of the reported patches, whereas Italy contributes only 10 patches and several regions are not represented. Therefore, the pooled accuracy should not be interpreted as geographically uniform European performance. Country- or ecological-region-level results and an appropriate macro-average would provide a more balanced assessment.In addition, the ten pixels sampled within each 640 m patch are spatially correlated and share the same ALS acquisition and processing characteristics. Validation should therefore also be summarized at the patch level, or uncertainty should be estimated using patches as the resampling unit. Confidence intervals are particularly important before interpreting the very small performance differences among the models as meaningful.
4. Temporal and phenological mismatch in the ALS validation
The map represents 2020, whereas the ALS observations span 2011–2023 and include both leaf-on and leaf-off acquisitions. The current filtering only reduces mixtures of flight dates within a patch; it does not address differences between the ALS acquisition and the target year. Consequently, the reported error includes forest growth, disturbance, and seasonal effects in addition to retrieval error.The authors should quantify how validation performance changes with temporal distance from 2020 and with acquisition season, and should more clearly state that the reported RMSE is not a purely contemporaneous mapping error.
5. Different spatial supports of GEDI and ICESat-2
GEDI samples are linked to predictors averaged over a 3 × 3 window, whereas ICESat-2 samples use a single 10 m pixel. Thus, the two training sources have different spatial supports when they are combined in the same model. It is also unclear whether wall-to-wall prediction uses the raw 10 m predictors or the same spatial averaging applied to the GEDI samples.This choice requires clearer justification and at least a simple sensitivity analysis comparing the current approach with a common spatial-support treatment. The manuscript should also distinguish between the clear benefit of adding ICESat-2 to the GEDI-only model in northern Europe and evidence for actual synergy between the two sensors. Relative to the ICESat-2-only model, the combined model improves the European RMSE only from 5.78 to 5.77 m and performs slightly worse in southern Europe.
6. Predictive uncertainty
The model predicts q0.05, q0.50, and q0.95, but the uncertainty estimates are not validated. The authors should demonstrate whether the nominal 90% interval achieves approximately 90% empirical coverage, including across different height ranges and geographic regions.Moreover, the public map repository appears to contain only q0.50 rasters. The q0.05 and q0.95 predictions, or the derived PI90/PIR layers, should be released if uncertainty is presented as part of the product. Otherwise, the related claims should be weakened. Equation 1 also defines prediction-interval width rather than the interval itself, and the treatment of PIR where q0.50 is zero or close to zero should be clarified.
7. Additional methodological information
Several processing steps need concise but more complete documentation. For Sentinel-1, the authors should specify whether γ⁰ was expressed in linear or dB scale, the applied radiometric and terrain corrections, the multitemporal speckle filter, and whether ascending and descending observations were combined or treated separately. These choices directly affect the definitions of VH–VV and RVI.The 100 × 100 km grid should be defined by its CRS and origin, and the cell-allocation procedure and random seed should be reported. The rationale and implementation of the ±1° coordinate perturbation also require clarification. Finally, training excludes slopes above 10%, whereas validation excludes slopes above 20% and the final map apparently covers all terrain. The authors should quantify the resulting excluded area and discuss the reliability of predictions over mountainous regions.
Minor comments
1. Clarify whether the ALS metrics were calculated for all valid pixels or only for pixels classified as woody vegetation.2. Clearly distinguish performance on the held-out GEDI/ICESat-2 test set from performance against the independent ALS data, particularly in Figure 7.
3. Correct the inconsistency between “S1S2-9B” in Table 1 and “S2S2-B9” in Figure 7.
4. Use consistent terminology for forest canopy height and woody-vegetation height, as the training and prediction domains also include shrubs and non-woody zero-height samples.
Citation: https://doi.org/10.5194/essd-2026-329-RC3 -
AC3: 'Reply on RC3', Wanda De Keersmaecker, 30 Oct 2026
This manuscript presents a 10 m canopy-height product for Europe derived from Sentinel-1 and Sentinel-2 data and trained using GEDI and ICESat-2 observations. The modelling framework itself is largely based on established approaches. However, I do not consider the limited methodological novelty to be a fundamental problem for ESSD. The main community value lies in the Europe-wide data product and, particularly, in the effort devoted to compiling and processing a geographically extensive independent ALS validation dataset. The reported performance is competitive, and the annual temporal compositing and filtering appear to produce a useful product.
I therefore consider the datasets potentially suitable for publication in ESSD. However, the present version does not yet meet the journal’s expectations for data completeness, reproducibility, uncertainty characterization, and transparent quality assessment. I recommend major revision.
Review of “Mapping forest canopy height over Europe by integrating Sentinel-1, Sentinel-2, GEDI, and ICESat-2 data”
We would like to thank the reviewer for his/her thorough review and for the constructive feedback.
Major comments
- Consistency of the released ALS dataset
The public validation dataset does not appear to match the dataset described in the manuscript. Supplementary Table S2 reports 3,695 patches across 18 countries, including 385 patches from Poland, whereas the Zenodo record contains 3,310 patches across 17 countries and does not include Poland. The difference is exactly 385 patches. Please reconcile this discrepancy and systematically verify that the manuscript, repository description, spatial index, and released files are mutually consistent.
The patches of Poland are indeed not included on Zenodo. We wanted to ensure that the data licence allows this and were awaiting confirmation. Although we may update the Zenodo record later, we added a remark about this in the manuscript under ‘code and data availability’ at lines 375-376:
“The patches over Poland are not included in this Zenodo release for licensing purposes.”
2. Reproducibility of the modelling workflowThe present description is insufficient to reproduce the model-derived product. Although the quantiles and general weighting approach are described, essential information such as the CatBoost version, complete hyperparameters, stopping procedure, random seeds, exact weighting implementation, and GEDI–ICESat-2 sample balance is missing. The final numbers of training, validation, and test samples for each sensor should also be reported.
At minimum, the manuscript and repository should provide all parameters, software versions, spatial split information, and processing details required to reproduce the results. Archiving version-controlled code and an environment file would be the clearest way to achieve this. CatBoost should also be properly cited.
We agree and added the information to the supplementary information S.5:
“The CatBoost model (CatBoostRegressor model; CatBoost version 1.2.5) was calibrated using a learning rate of 0.1, loss_function equal to 'MultiQuantile:alpha=0.05,0.5,0.95', iterations equal to 30000, 20 early_stopping_rounds, and a random_state equal to 42. The distribution of the GEDI and ICESat-2 samples is shown in figure S2. In total, 362508, 50614, and 39807 samples were available from GEDI and 870309, 117669, and 101387 samples from ICESat-2 for training, validation, and testing, respectively. 491717, 67308, and 55869 samples in the training, validation and test dataset were sampled outside woody areas. The sampling grid had an upper left corner with coordinate 2600 000 (x), 6600 000 (y) in EPSG 3035.”
3. ALS validation design and geographic representativeness
The ALS dataset is extensive but strongly concentrated in a few countries. Spain, Norway, Finland, and Sweden account for approximately 57% of the reported patches, whereas Italy contributes only 10 patches and several regions are not represented. Therefore, the pooled accuracy should not be interpreted as geographically uniform European performance. Country- or ecological-region-level results and an appropriate macro-average would provide a more balanced assessment.In addition, the ten pixels sampled within each 640 m patch are spatially correlated and share the same ALS acquisition and processing characteristics. Validation should therefore also be summarized at the patch level, or uncertainty should be estimated using patches as the resampling unit. Confidence intervals are particularly important before interpreting the very small performance differences among the models as meaningful.
The reviewer is correct; the ALS dataset is not equally covering the whole area as only freely available data is collected. This is stated in the text (line 236-237), illustrated in Figure 1 and discussed in lines 343-345. To our knowledge, it still represents the most extensive ALS dataset collected over Europe. In our opinion, the main added value of the ALS dataset is that it provides an independent assessment, closer to in-situ measurement, compared to space-borne LiDAR. With the ALS dataset, our objective is indeed to go beyond the evaluation based on the test set, usually used in other studies. The evaluation with the test set is now also added in supplementary information S.6. It is aligned with the evaluation with the ALS dataset.
Regarding the sampling design, we used a two-stage sampling approach. First, ALS patches were selected using stratified random sampling. Second, ten pixels were randomly sampled within each selected patch. The accuracy metrics were calculated at the 10 m spatial resolution of the canopy-height products, preserving the spatial support of the maps being evaluated. We however agree that pixels sampled within the same 640 m patch may be spatially correlated and share acquisition and processing characteristics. All models were evaluated using the same reference pixels, allowing for a paired comparison under the same validation conditions. This does not remove the potential effect of on the uncertainty of the accuracy metrics, particularly when interpreting very small differences in performance between models. We added this limitation to the text at lines 349-351.“Finally, uneven geographic coverage and spatial dependence between pixels may affect accuracy estimates, so small differences between models should be interpreted with caution.”
4. Temporal and phenological mismatch in the ALS validation
The map represents 2020, whereas the ALS observations span 2011–2023 and include both leaf-on and leaf-off acquisitions. The current filtering only reduces mixtures of flight dates within a patch; it does not address differences between the ALS acquisition and the target year. Consequently, the reported error includes forest growth, disturbance, and seasonal effects in addition to retrieval error.The authors should quantify how validation performance changes with temporal distance from 2020 and with acquisition season, and should more clearly state that the reported RMSE is not a purely contemporaneous mapping error.
We agree that the heterogeneity of the dataset is a limitation. Since this is relevant for the validation and for other potential users of the dataset, we added a paragraph to the discussion to describe this (lines 343-349):
“While we believe that the ALS dataset provides valuable information on canopy structure, several limitations should be considered when using the data. Firstly, the ALS data across Europe is heterogeneous, having been collected using different instruments and on various dates. For instance, some data from Poland was collected in 2011, whereas parts of France were surveyed in 2023. In addition, the day of year varies across the dataset. This discrepancy in the ALS flight dates may cause relatively small discrepancies due to growth or phenology and larger ones due to tree loss. Secondly, also the point density of the ALS measurements differs across the dataset. Lower point densities reduce the capacity to capture the fine-scale spatial positioning of vegetation components, potentially leading to less reliable vegetation structure metrics.”
It’s an interesting suggestion to perform such analysis, but since the year and day of year is spatially aggregated, the result would be a mixture of spatial accuracy, temporal disagreement, and number of samples. To further strengthen the validation of the datasets, we added the validation over the test set to the supplementary information (S6), which points to a similar conclusion.
“The validation of the canopy height estimates from the canopy height model and other datasets available over Europe is shown in Figure S3. The results were obtained using observations inside and outside woody areas.”
5. Different spatial supports of GEDI and ICESat-2
GEDI samples are linked to predictors averaged over a 3 × 3 window, whereas ICESat-2 samples use a single 10 m pixel. Thus, the two training sources have different spatial supports when they are combined in the same model. It is also unclear whether wall-to-wall prediction uses the raw 10 m predictors or the same spatial averaging applied to the GEDI samples.
This choice requires clearer justification and at least a simple sensitivity analysis comparing the current approach with a common spatial-support treatment. The manuscript should also distinguish between the clear benefit of adding ICESat-2 to the GEDI-only model in northern Europe and evidence for actual synergy between the two sensors. Relative to the ICESat-2-only model, the combined model improves the European RMSE only from 5.78 to 5.77 m and performs slightly worse in southern Europe.The different spatial windows for the predictors were selected to better represent the footprint of the spaceborne lidar datasets. When performing the wall-to-wall map prediction we used the raw 10m predictors. We added a clarification to the text at lines 162-163:
“… . Different spatial windows for the predictors were selected to represent the footprint of the spaceborne lidar datasets. …”
Although a sensitivity analysis would be interesting to perform, we believe that this is outside the scope of this manuscript.
6. Predictive uncertainty
The model predicts q0.05, q0.50, and q0.95, but the uncertainty estimates are not validated. The authors should demonstrate whether the nominal 90% interval achieves approximately 90% empirical coverage, including across different height ranges and geographic regions.Moreover, the public map repository appears to contain only q0.50 rasters. The q0.05 and q0.95 predictions, or the derived PI90/PIR layers, should be released if uncertainty is presented as part of the product. Otherwise, the related claims should be weakened. Equation 1 also defines prediction-interval width rather than the interval itself, and the treatment of PIR where q0.50 is zero or close to zero should be clarified.
We agree and weakened the claims in the text. We called the metrics indicators of uncertainty instead of uncertainty estimates and added the following sentence at lines 190-191:
“Since these uncertainty metrics were not validated, they only serve as an indicator of uncertainty, but not necessarily represent the true uncertainty.”
The PIR metric was set to nodata over observations with q0.50 equal to zero.
7. Additional methodological information
Several processing steps need concise but more complete documentation. For Sentinel-1, the authors should specify whether γ⁰ was expressed in linear or dB scale, the applied radiometric and terrain corrections, the multitemporal speckle filter, and whether ascending and descending observations were combined or treated separately. These choices directly affect the definitions of VH–VV and RVI.
The 100 × 100 km grid should be defined by its CRS and origin, and the cell-allocation procedure and random seed should be reported. The rationale and implementation of the ±1° coordinate perturbation also require clarification. Finally, training excludes slopes above 10%, whereas validation excludes slopes above 20% and the final map apparently covers all terrain. The authors should quantify the resulting excluded area and discuss the reliability of predictions over mountainous regions.
We added more details to the description of the Sentinel-1 processing at lines 133-143:
“The Sentinel-1 features were derived in a similar fashion to the Sentinel-2 features. The Sentinel-1 Ground Range Detected (GRD) dual-polarisation (VV, VH) scenes were first corrected to Gamma0 backscatter and geocoded to the 20m Sentinel-2 grid using a geoid-corrected (EGM2008) DEM. Radiometric calibration with thermal-noise (NESZ) removal, 2×2 multilooking, terrain radiometric terrain correction by DEM-based pixel-area normalisation, and conversion to γ⁰ using the local incidence angle were applied. Layover or shadow pixels were masked. Maximum one observation per day and per tile was processed to reduce data redundancy and ascending and descending passes were combined into a single time series. A multitemporal speckle filter (Quegan-type, 7×7 Gamma-MAP kernel, ENL = 3) was then applied to reduce noise, after which the time series were aggregated into ten daily composites using a 20-day moving-median composites and linear gap interpolation. All filtering and compositing was done in linear power scale. The composites were then converted to dB (10·log10), on which the temporal statistics were computed. The percentiles and interquartile range were computed on the Sentinel-1 VV and VH backscatter, the difference between the VH and VV backscatter, and the radar vegetation index (RVI) time series. The RVI was computed from the linear composites.”
The CRS and origin of the grid and the sampling strategy was added to supplementary information S5.
We added the rationale for the noise on the coordinates at lines 175-178:
“The purpose of the latitude and longitude features is to allow the model to be aware of the approximate location within Europe. Localizing information gives the model some biogeographical context and might therefore help the model to better interpret the signal. Yet, since the location is unique for each sample, there is some risk to overfit on this feature. To avoid this, we added noise to the latitude and longitude that was randomly sampled between -1 and 1 degrees.”
We discussed the exclusion of training data over areas with steep slopes at lines 363-365:
“… Finally, given the sensitivity of LiDAR data to terrain slope, steep slopes were excluded both from the training and validation dataset. As such, it is unknown how accurate the canopy height estimates are over these areas.”
Minor comments
1. Clarify whether the ALS metrics were calculated for all valid pixels or only for pixels classified as woody vegetation.
The ALS-based metrics were computed from both woody and non-woody pixels
2. Clearly distinguish performance on the held-out GEDI/ICESat-2 test set from performance against the independent ALS data, particularly in Figure 7
We added the information to the caption of Figure 7
3. Correct the inconsistency between “S1S2-9B” in Table 1 and “S2S2-B9” in Figure 7.
We corrected the mistake.
4. Use consistent terminology for forest canopy height and woody-vegetation height, as the training and prediction domains also include shrubs and non-woody zero-height samples.
We used the terminology woody vegetation in the title and the manuscript. In the introduction, we kept the term forest to situate the work in a broader context.
Citation: https://doi.org/10.5194/essd-2026-329-AC3
-
AC3: 'Reply on RC3', Wanda De Keersmaecker, 30 Oct 2026
Data sets
European canopy height map W. De Keersmaecker et al. https://doi.org/10.5281/zenodo.13324731
ALS-based canopy height across Europe L. Bertels et al. https://doi.org/10.5281/zenodo.18471620
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 578 | 325 | 143 | 1,046 | 207 | 176 | 131 |
- HTML: 578
- PDF: 325
- XML: 143
- Total: 1,046
- Supplement: 207
- BibTeX: 176
- EndNote: 131
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
The authors present a study aimed at producing a 10 m forest canopy height map for Europe by integrating Sentinel-1, Sentinel-2, GEDI and ICESat-2 observations. Overall, this is a good study that complements the growing landscape of continental and global canopy height products. Whilst the overall modelling framework is largely based on established approaches, the work provides an important contribution by extending training to northern Europe through the integration of ICESat-2 data and by delivering an openly available European canopy height dataset.
In my opinion, one of the strongest aspects of the manuscript is the validation strategy. The authors assembled an extensive independent validation dataset from heterogeneous airborne LiDAR datasets collected across multiple European countries. Such a large-scale independent validation effort is relatively uncommon and substantially increases confidence in both the reported accuracy and the resulting data product.
The manuscript is generally well written, the methodology is clearly described, and the data product will likely be valuable for a wide range of ecological and forestry applications. My comments below are intended to further strengthen the manuscript.
Comments:
1. The independent ALS validation dataset is one of the principal strengths of this study. However, the ALS acquisitions span multiple years, while the canopy height map represents conditions in 2020. Although the manuscript acknowledges this limitation, I encourage the authors to discuss more explicitly how temporal mismatches may influence the reported validation statistics.
2. Clarify the added value of this dataset. The manuscript compares the proposed product with several existing canopy height datasets and demonstrates modest improvements in validation statistics. While these improvements are encouraging, I suggest that the Discussion better emphasise the broader advantages of the dataset. In particular, the integration of ICESat-2 substantially improves training coverage in northern Europe, the extensive independent validation dataset increases confidence in the results, and the provision of uncertainty estimates enhances the usefulness of the product. These aspects arguably represent the primary contribution of the study and deserve greater emphasis than the relatively small differences in RMSE.
Minor comments:
In the Abstract and the Introduction, "spatial explicit" should be corrected to "spatially explicit"
Please clarify whether the reported RMSE and MAE are computed over all validation pixels or only over pixels classified as woody vegetation.
Table 1 refers to the combined Sentinel-1/Sentinel-2 model as "S1S2-9B", whereas Figure 7 uses "S2S2-B9". This appears to be a typographical inconsistency and should be corrected.
The manuscript occasionally uses the terms height estimate, height prediction, and canopy height map interchangeably. Using consistent terminology throughout would improve readability.
The choice of adding random noise of ±1° to the latitude and longitude variables, together with the decision to saturate the sample weights above 35 m canopy height, would benefit from a brief justification.
A concise table summarising the main characteristics of this product and the existing canopy height datasets used for comparison (e.g. training data, spatial resolution, spatial extent, independent validation, and uncertainty estimates) would provide useful context for readers.