acceptodds
Under review as a conference paper at ICLR 2027

ECHOSAT: Estimating Canopy Height Over Space and Time

Abstract

Forest monitoring is critical for climate change mitigation. However, existing global tree height maps only provide static snapshots and do not capture temporal forest dynamics, which are essential for accurate carbon accounting. We introduce ECHOSAT, a global and temporally consistent tree height map at 10\meter resolution spanning multiple years, which accurately quantifies tree growth and disturbances over time. To this end, we resort to multi-sensor satellite data to train a vision transformer model, which performs pixel-level temporal regression. A self-supervised growth loss regularizes the predictions to follow growth curves that are in line with natural tree development, including gradual height increases over time, but also abrupt declines due to forest loss events such as fires. Our evaluation further shows that our model improves state-of-the-art accuracies in the context of single-year predictions. ECHOSAT can be accessed here: https://echosatanonymous.projects.earthengine.app/view/echosat-temporal.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
climate science, computer vision, canopy height, remote sensing, application

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_LjPLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1(−1)
Read review
Paper summary

ECHOSAT introduces a spatio-temporal framework using multi-source satellite remote sensing data. The framework is designed to generate annual global 10m tree height maps for the period 2018–2024, while capturing height variations driven by forest growth and disturbances.

The authors developed a Swin Transformer-based spatio-temporal U-Net architecture, integrating data from Sentinel-1/2, ALOS PALSAR-2, and TanDEM-X, while utilizing GEDI LiDAR measurements as training labels. A key methodological contribution is the introduction of "Growth Loss"—a technique that generates pseudo-labels by applying piecewise linear regression (constrained by growth rates) to initial tree-height predictions. This approach enhances the temporal consistency of annual predictions while still allowing for abrupt height changes caused by forest disturbances.

The authors have compared ECHOSAT with 4 existing global canopy-height products and further assessed prediction accuracy using independent airborne LiDAR data from the NEON network. Testing against 2020 GEDI data revealed that ECHOSAT achieved a Mean Absolute Error (MAE) of 5.85 m, outperforming the best-performing baseline (6.85 m). The authors also evaluated the impact of Growth Loss on temporal consistency through regional case studies, analyses of annual height-change distributions, and ablation experiments.

Critical strengths

The GEDI data is sparse in both space and time, this paper propose one solution, which name is "Growth Loss" function that incorporates ecological constraints, thereby improving the temporal consistency of predictions. This is valuable for large-scale forest monitoring. The underlying concept—integrating physical constraints into model design—serves as an effective approach for handling sparse remote sensing data.

Critical weaknesses

I have listed 2 weaknesses:

  1. Lack of independent quantitative validation for year-to-year changes in canopy height.

The experiments in the paper primarily validate the prediction accuracy of canopy height for a single year. Regarding temporal dynamics, the authors rely on evaluating predicted growth rate distributions, qualitative case studies, and temporal consistency analyses; however, there is a notable absence of independent quantitative validation using diverse LiDAR datasets (ULS/MLS/TLS) or ground-based forest plot observations. Given that annual growth increments are typically much smaller than the prediction error for any single year, demonstrating that the results are "smoother" or consistent with predefined growth patterns does not prove that the model has accurately captured actual changes in forest height. Consequently, the paper’s most significant contribution—the modeling of temporal dynamics—remains insufficiently validated.

  1. A lack of comparative experiments to clearly demonstrate the method's contributions.

The proposed architecture is a combination of existing components—Swin Transformer and U-Net—while the "Growth Loss" is based on piecewise linear regression with predefined constraints. The current experiments compare the model primarily against static canopy height products; there is a lack of direct benchmarking against other single-year or time-series models using identical input data and training conditions. Furthermore, ablation studies indicate that single-year prediction accuracy remains virtually the same whether or not the "Growth Loss" is applied. As a result, it is unclear to what extent the performance gains stem from the proposed spatiotemporal modeling approach itself, which complicates an assessment of the method's true innovation and practical contribution.

Questions for authors

Q1. Question on similar products

The Monitoring 10 m product by Kanop, in its 2024 release, provides global annual canopy height estimates (covering both "canopy height" and "tree height") at a 10 m resolution, dating back to 2017. Could the authors further clarify the primary differences between ECHOSAT and this existing product?

Q2. Disentangling "Growth Loss" from the Contributions of the Spatiotemporal Model Architecture

Appendix Table 4 indicates that there is virtually no difference in single-year prediction accuracy between a model incorporating "Growth Loss" and one without fine-tuning. While "Growth Loss" does improve temporal consistency, this improvement appears to stem primarily from the predefined growth constraints. Do the authors have control experiments—such as a comparison between independent single-year prediction models and a joint spatiotemporal model—conducted under identical input data and experimental conditions to demonstrate the performance gains attributable specifically to the spatiotemporal modeling framework? Furthermore, is there independent evidence to ensure that "Growth Loss" does not lead to the over-smoothing of actual height changes, particularly in the case of smaller-scale forest degradation events? As Growth Loss is key contribution of this work, I want to know these additional results.

Q3. Dependence of model predictions on future observations

Based on the model architecture, satellite imagery from 2018–2024 is fed into the model collectively to generate canopy height predictions for all seven years simultaneously. For a given year—for instance, 2019—does the model utilize information from imagery acquired in 2020–2024? If so, these annual height maps effectively constitute a "retrospective reconstruction" that relies on future observations, rather than independent estimates that could have been generated at the time for each respective year. Could the authors clarify this point, and explain its implications for the paper’s stated application of annual forest monitoring?

5 comments

We thank the reviewer for the feedback. We would like to point out that the GrowthLoss is not our only contribution: The main asset of our paper is the ECHOSAT map itself in our optinion.

Below, we have tried to address the four concerns raised by you.

Temporal change validation

We fully agree with your concern. Validation of the temporal changes is important for downstream users, however globally stratified repeated LiDAR observations are needed for such a validation. To the best of our knowledge, no such dataset exists. As you rightly point out, growth increments are typically much smaller than the prediction error for an individual year. This limits the usefulness of competing approaches that treat each year independently and was a key motivation for developing GrowthLoss. Nevertheless, although it is only for a local region (USA), we have done some analysis using the repeated NEON sites, which you can find in our response to reviewer HuEN.

Contribution of GrowthLoss and model architecture

We believe it is important to distinguish three aspects of this comment: the contribution of the GrowthLoss, the contribution of the model architecture, and the extent to which GrowthLoss may oversmooth the predicted heights.

  1. Contribution of the GrowthLoss: You are correct that Table 4 does not show any differences, which is consistent with the intended behavior of the loss. We developed the GrowthLoss to obtain more realistic temporal tree heights, while not degrading the single-year performance. The reference head and the prediction head share exactly the same backbone architecture and data. Table 3 shows that the the GrowthLoss leads to fewer spurious height decreases, and Figure 9 shows that the piecewise design matters to deal with disturbances.

  2. We are not able to provide any quantitative evidence for the contribution of the model architecture in this short (rebuttal) time frame, as training the global model takes substantial computational efforts (8 NVIDIA H200s for 1 week). During the development we started with single-year baselines, such as the one provided by Pauls et al. (2025) or Kacic et al. (2023), however we noticed that there are still unrealistic height fluctuations. To be able to use a loss function across all years (like with the GrowthLoss), our model needs to be able to process all years at once, and we developed the spatiotemporal architecture accordingly. Nevertheless, we evaluated the model on a sample of our validation set, when certain inputs are zeroed out:

ConfigurationMAEMSERMSEMAPER²R²_allrr_all
Ours only S27.045143.15311.96534.5260.0710.4770.5220.709
Ours S2+S16.653134.02711.57732.4180.1310.5110.5420.731
Ours S2+S1+PALSAR26.613134.87911.61431.9640.1250.5160.5490.745
Ours S2+S1+PALSAR2+DEM5.823118.90310.90429.1510.2290.5730.5880.771
Ours all5.788117.96110.86129.2930.2350.5770.5910.771
  1. Our disturbance indicator has three conditions (height decreases by more than 50%, height decreases by more than 4 meters and height is below 10m in one of the next two years) and is for example not able to detect small-scale disturbances. These are then fitted with a non-negative slope, hence being oversmoothed. That means that ECHOSAT is not able to detect selective logging and thinning for example. Table 3 shows that gradual decreases drop from 23% to 13% when using the GrowthLoss. As already stated in Section 5.1, some of the gradual decreases are real and missed by our map.

Thank you for your comment. We will update the paper to emphasize that the accuracy gain comes from the full pipeline (data, model archticture, training, preprocessing, ...)

[1] Pauls et al. (2025): Capturing Temporal Dynamics in Large-Scale Canopy Tree Height Estimation. Proceedings of the 42nd International Conference on Machine Learning

[2] Kacic et al. (2023): Forest structure characterization in Germany: novel products and analysis based on GEDI, Sentinel-1 and Sentinel-2 data. Remote Sensing, 15(8), 1969.

Retrospective vs live product

This is correct. The prediction for the year 2019 is based on all years, including 2020-2024. Our goal, guided by input from domain experts, is to produce the best canopy height map possible, which also means including every data point that could increase the accuracy. Nevertheless, we would like to provide results for causal masking and report the results on the same set as for Table 1. As we can see, the results do not differ by much, sometimes even improving. Further, one can run the pipeline every year to get an updated version.

MethodMAEMSERMSEMAPER²R²_allrr_all
Causal masking5.802118.28410.87629.2850.2330.5760.5920.772
Ours5.854118.07410.86630.2010.2340.5770.5880.770

Comparison to Monitoring 10m product by Kanop

Kanop is a commercial platform based on proprietary models. To the best of our knowledge, its methodology, training data and validation are not publicly documented or peer-reviewed, and its maps are only available through paid access. We therefore could not include it in our benchmark, which uses openly available global products. Methodologically, it is unclear whether Kanop models temporal dynamics or applies per-year predictions. ECHOSAT learns them directly via its spatio-temporal architecture and growth loss, without post-processing. ECHOSAT is also fully documented and will beopenly released (code: Apache 2.0, maps: CC-BY 4.0).

If you would like further clarifications, please do not hesitate to ask.

Dear Authors,

Many thanks for your feedback, I will provide responses based on your response.

  1. Temporal change validation

I have read the response to reviewer HuEN; the results highlight two significant issues:

  • Severe Underestimation of Change Magnitude: ECHOSAT’s estimates of tree height change are compressed by a factor 4–8, this is a strong number.
  • Inadequate Detection of Declines: In where ALS data indicates a decrease in tree height, the median change predicted by ECHOSAT remains positive (this is realted to no negative setting of the model).

This new validation reveals a systematic bias in the magnitude of change; the model has no reliable capability for quantifying tree height changes at the pixel level.

Also, the authors do not specify how many sites are covered by the NEON test.

While I acknowledge the ongoing multi-year LiDAR validation datasets in global scale, NEON is a good start.

UCL, Ghent University, and UMR AMAP in France have made ALS/TLS data publicly available for regions in South America, Africa, and Southeast Asia—for instance, through projects like Weighing Trees with Lasers and ForestScan. I suggest to do test in the authors' study for validation.

  1. Contribution of GrowthLoss and model architecture

I understand that the "Growth Loss" component is intended to improve temporal consistency; however, temporal consistency is not synonymous with temporal accuracy.

I think this work have not demonstrated that the "Growth Loss" approach recovers actual forest dynamics more accurately than a model without such fine-tuning. In fact, it may smooth out genuine declines in tree height that fall below the threshold for disturbance detection.

Maybe using NEON ALS and other LiDAR data to compare the prediction errors of tree height change—as well as the capacity for disturbance detection—both before and after applying the "Growth Loss" fine-tuning can help to calrify the contribution of the model.

Reviewer_oWMfOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

ECHOSAT estimates and performs inference to produce annual canopy height at 10m res for 2018–2024 using imagery from multiple sensors as input and a temporal Swin-Unet. A second head learns pseudo-labels by fitting growth-constrained piecewise-linear trajectories to a frozen reference head.

Critical strengths
  • the paper provides the first 10m global multitemporal canopy height map
  • the paper and supplementary material provides significant details for the implementation and run
  • the dataset and code are clearly available during review and the paper doesn't read like LLM language
  • the gains are clear over GEDI and NEON airborne lidar datasets
  • the temporal regularization and consistency is an interesting methodology that other global multitemporal datasets like Fields of the World (FTW) could apply
Critical weaknesses
  • the paper does not cite or compare to Pauls et al (ArXiv 2025) published at ICML 2026 [1] which produces a 10 m temporal canopy-height maps for Europe using Sentinel-1/2 and GEDI. This is probably the largest weakness of the paper. They compare to Pauls et al (ICML 2024) though but this is an older version.
  • the limitations state that the data product and model cannot be evaluated in areas without ground truth. This dataset is similar in spirit to the global Fields of the World (FTW) [2] which also suffered from this problem and they came up with a solution to provide a model confidence layer to identify OOD areas. It would be useful for ECHOSAT to provide a similar output layer for downstream users

[1] https://icml.cc/virtual/2025/poster/43876 [2] https://arxiv.org/abs/2605.11055

Questions for authors

While [1] doesn't provide global canopy height it does provide 10m multi-temporal canopy height over Europe which weakens the novelty of the paper which reduces to it just being global. Then it does not cite or compare to this dataset or paper methodology while it's probably the closest related work and published to a top conference (ICML 2026). How does ECHOSAT compare to this? I opted for weak accept but happy to increase my score if this comparison is done.

[1] https://icml.cc/virtual/2025/poster/43876

2 comments

We thank the reviewer for the constructive feedback and adress the two concerns below.

Comparison to Pauls et al. (2025)

Although Pauls et al. (2025) is the newer version, it only covers Europe and our goal was to compare against other global maps, as we argue that forest monitoring should be a global effort. Nevertheless, to address your concern, we have conducted the comparison using all validation samples in Europe and report the results below.

MAEMSERMSEMAPER2R2_allrr_all
Potapov et al.10.367231.51715.21658.056-0.5070.0870.3710.587
Lang et al.7.743149.42712.22446.5570.0270.3690.4120.629
Pauls et al. (2024)7.481163.0112.76836.07-0.0610.3660.4030.636
Tolan et al.12.474286.92216.93969.23-0.868-0.1220.3770.566
Ours6.813152.49812.34932.8930.0070.4060.4540.665
Pauls et al. (2025)6.655154.0512.41231.718-0.0030.3980.4590.669

The table shows that our map and the map by Pauls et al. (2025) are very similar: MAE, MAPE, pearson_r and pearson_r_all results are better for Pauls et al. (2025), whereas MSE, RMSE, R2 and R2_all are better for our map. We further compared the predicted change for consecutive years and added them to the appendix. The newly added Figure 13 compares the predicted change between our map, Pauls et al. (2025) and Turubanova et al. It shows that Pauls et al. (2025) shows a lot of change for pixels, where our map and Turubanova et al. (2025) do not predict no change (vertical line near x=0 in a) and b)), however this pattern is not present in the comparison between our map and Turubanova et al. Figure 14 shows the density plot/histogram of predicted year-to-year change of all three maps. It can be seen that Pauls et al. (2025) exhibits a lot of change, that both other maps do not report, as expected. Note that Pauls et al. (2025) already uses a quadratic spline smoothing (see under "Post-Processing", p. 4), but has a lot of small fluctuations.

[1] Pauls et al. (2025): Capturing Temporal Dynamics in Large-Scale Canopy Tree Height Estimation. Proceedings of the 42nd International Conference on Machine Learning

[2] Turubanova et al. (2025): Tree canopy extent and height change in Europe, 2001–2021, quantified using Landsat data archive. Remote Sensing of Environment, 298, 113797.

Out-of-Distribution Areas

That is good point. Although we deal with a regression task, and not with an object (field) detection task, we see the benefit of such an additional layer. In our case, however, the out-of-distribution area is easy to describe as it solely relies on the availability of GEDI data. We used all available GEDI labels and therefore covered the entire Earth's land mass between 51.6°N and 51.6°S. We will add this to the limitations so users are aware of this.

Thank you again for your comments. Please let us know if you have any further concerns.

Thank you for the additional comparisons to the points I requested. While I agree with other reviewers that there are limitations to validation due to the availability of existing global and multitemporal ground truth, I still believe this is a valuable contribution to the geospatial research and industry community and therefore will be increasing my score as I said I would in my review.

Reviewer_HuENOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ECHOSAT, a method for global, temporally consistent tree height mapping, and presents annual maps at 10 m resolution for 2018–2024. It combines Sentinel-2 optical imagery, Sentinel-1 and ALOS PALSAR-2 radar data, and TanDEM-X elevation and forest classification data with sparse GEDI supervision. Its core architecture, Temporal-Swin-Unet, adapts Video Swin Transformer and Swin-Unet to jointly predict annual heights from monthly observations. The backbone and reference head are first trained with a Huber loss. They are then frozen while a separate prediction head is trained with the proposed Growth Loss against pseudo-labels generated through disturbance detection and constrained piecewise linear regression.

The paper’s three stated contributions are the global seven-year maps, a Growth Loss for spatially and temporally sparse supervision, and the ability to capture forest growth and disturbances without post-processing. Experiments include year-wise and stratified evaluations, comparisons with four existing maps, independent NEON airborne laser scanning validation, and ablations. The results show improved single-year accuracy across the reported metrics and reduced temporal fluctuations, with qualitative examples showing that abrupt disturbances are retained.

Critical strengths

1. A substantial contribution to global temporal tree height mapping. Combining multi-sensor data, joint temporal modeling, and growth constraints makes a lot of sense given how difficult it is to learn from sparse height labels. Also, generating global 10 m maps across seven years is a massive effort and will be a useful resource for studying forest dynamics.

2. Consistent single-year accuracy gains. ECHOSAT outperforms all four comparison maps across the reported metrics on held-out GEDI measurements, and independent NEON ALS observations confirm the ranking (Tables 1 and 8). The stratified analyses help characterize performance across heights and regions. Table 4 further shows that fine-tuning with Growth Loss largely preserves single-year accuracy.

Critical weaknesses

1. The accuracy of predicted changes is not directly validated. The strongest quantitative evidence concerns single-year heights. Section 4.2 and Table 3 assess whether predictions exhibit plausible growth patterns and fewer fluctuations, but do not establish accurate growth increments or disturbance timing. Since the Growth Loss explicitly encourages these patterns, these results demonstrate that the constraint works as intended, rather than independently confirming recovery of actual forest dynamics. The paper’s third stated contribution therefore remains insufficiently supported.

2. The additional benefit and limitations of the Growth Loss remain unclear. Table 4 supports preservation of single-year accuracy but does not compare change-estimation accuracy across the training variants. Figure 9 illustrates temporal behavior for only five neighboring pixels around one event. Genuine losses that fail to meet the disturbance criteria may also be smoothed. It is therefore unclear whether the proposed loss consistently improves on simpler temporal constraints while preserving real changes.

Questions for authors

**1. Can repeated airborne LiDAR or forest plot measurements validate the predicted growth increments and disturbance-related losses in representative areas, including event timing where possible? Please clarify whether agreement holds at the pixel level, after spatial aggregation, or over multi-year intervals. This would establish the scope of the claimed temporal accuracy.

**2. On these independent reference samples, how do the three settings in Table 4 compare in estimating growth and recovering disturbances? Keeping the two fine-tuning settings otherwise comparable would help isolate the loss’s contribution and determine whether the advantage illustrated in Figure 9 extends beyond that example.

**3. Do the main findings remain stable under reasonable changes to the disturbance thresholds and maximum growth slope? A focused analysis in a few representative areas, including genuine losses near or below the detection thresholds, would help determine whether reduced fluctuations come at the cost of missed changes.

1 comment

We thank the reviewer for acknowledging the substantial contribution to global temporal tree height mapping. Below we address your raised concerns and questions.

Accuracy of predicted changes

We acknowledge that validation of the predicted changes is not strongly supported by evidence. As we do not have a global dataset of tree heigth change, which we can use for validation, our strategy focused more on qualitative assessmenets and consistency checks. To address your concerns, we can provide a local analysis of the predicted changes. We haved used ALS data from the NEON campaign in the USA: Here, ee took every NEON site with two ALS acquisitions from the same sensor, either 2 or 3 years apart, and compared the ECHOSAT height change against the ALS height change.

2-year difference

ALS bin (m)nmeanq5medianq95
(-3.0,-2.0)336194-0.042-0.6160.0280.687
(-2.0,-1.0)7688380.005-0.5930.0370.765
(-1.0,-0.5)10622080.052-0.4840.0410.745
(-0.5,0.5)192181910.078-0.220.0380.566
(0.5,1.0)30335920.193-0.2940.1280.913
(1.0,2.0)16315310.335-0.3160.2161.46
(2.0,3.0)3344740.399-0.3210.2321.761

3-year difference

ALS bin (m)nmeanq5medianq95
(-3.0,-2.0)290250-0.162-1.360.0260.782
(-2.0,-1.0)662975-0.069-0.9520.0330.754
(-1.0,-0.5)9251950.013-0.5990.0340.691
(-0.5,0.5)162158490.056-0.2250.0320.521
(0.5,1.0)30313850.095-0.3190.0640.68
(1.0,2.0)18575340.18-0.310.1220.949
(2.0,3.0)5353970.297-0.2520.171.287

The following observations can be made:

  1. The model responds to the true magnitude of change, not only to its sign. The mean ECHOSAT change is monotone w.r.t. to the ALS change across all bins in both settings, and it separates weak from strong change. Visual inspection supports this: different forest types receive visibly different growth magnitudes, with fast-growing plantations showing clearly stronger increments than natural forests. The model is not producing a flat, constant growth offset.
  2. The amplitude is systematically compressed: Taking bin midpoints as the reference, the predicted increments are roughly 4-6 times too small at 2 years and about 8 times too small at 3 years. The median is positive in every bin, including bins where the ALS clearly measures height loss. At the pixel level the model is biased towards growth and is not able to reproduce gradual decline.
  3. The spread is large relative to the signal (see q5/q95).

It is worth noting that one cannot infer from this experiment whether the compression is an inherent limitation of ECHOSAT or if it refelects the constraints of sparse GEDI supervision (a single temporal GEDI observation for sparse pixels). Because of this, we have weakened the temporal claim througut the entire paper. E.g., Contribution 3 will be:

"Third, we demonstrate that our model captures abrupt disturbances and that its aggregate growth signal correlates with independent repeat-ALS measurements, while systematically underestimating absolute growth rates. At the same time, our model outperforms existing approaches in single-year evaluations."

Finetuning setting

Our GrowthLoss is able to detect disturbances, once they exceed the thresholds specified in the disturbance indicator. The disturbance indicator is however not able to detect gradual decreases and these are hence fitted with a single non-negative slope, i.e. oversmoothed. The example provided in Figure 9 is representative if the disturbance is a clear cut. In a setting where there is a stable forest, both fine-tuning variants are similar and smoother than the non-fine-tuned version. However, when the disturbance is only a drop below the indicator threshold, the not-fine-tuned version is able to detect it, but both fine-tuned versions are not. One can therefore say that the GrowthLoss oversmooths some scenarios to improve on others. We would argue that the prevalence of full clear-cuts and stable forest is much higher than partial cut, therefore this trade-off is acceptable.

Disturbance thresholding

We have conducted various experiments w.r.t the disturbance indicator and the specific disturbance thresholds. As we do not have high-quality disturbance data that are stratified globally, we have tried to counteract this limiation by talking to field experts and by visually testing the results at different known disturbance events. Therefore we cannot provide any quantitative results on this at the time of writing. However, during the develoment and application of our models, we noticed that changing most thresholds slightly did not lead to severe differences. The only parameter that had a greater impact was the "50% decrease" in the disturbance indicator that we, therefore, discussed in more details. As stated in our limitations (Section 5.1), the ECHOSAT maps are not able to detect selective logging and thinning for example.

Please let us know if you have further questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.