ECHOSAT: Estimating Canopy Height Over Space And Time

Jan Pauls; Karsten Schrödter; Sven Ligensa; Martin Schwartz; Berkant Turan; Max Zimmer; Sassan Saatchi; Sebastian Pokutta

cs.CVFeb 24, 2026

ECHOSAT: Estimating Canopy Height Over Space And Time

Authors: Jan Pauls, Karsten Schrödter, Sven Ligensa, Martin Schwartz, Berkant Turan, Max Zimmer, Sassan Saatchi, Sebastian Pokutta,

Organizations: Department of Information Systems, University of Münster, Germany · Laboratoire des Sciences du Climat et de l’Environnement (LSCE), France · Zuse Institute Berlin (ZIB), Germany · Jet Propulsion Laboratory (JPL), California Institute of Technology, USA ·

Abstract

Forest monitoring is critical for climate change mitigation. However, existing global tree height maps provide only 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 m resolution spanning multiple years. To this end, we resort to multi-sensor satellite data to train a specialized 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 experimental evaluation shows that our model improves state-of-the-art accuracies in the context of single-year predictions. We also provide the first global-scale height map that accurately quantifies tree growth and disturbances over time. We expect ECHOSAT to advance global efforts in carbon monitoring and disturbance assessment. The maps can be accessed at https://github.com/ai4forest/echosat.

Figures & tables

Figure 1: Qualitative comparison across three geographically diverse locations. The first column shows Google Maps imagery for spatial context, while subsequent columns display predicted tree heights ( $0\text{\,}\mathrm{m}$ $-$ $30\text{\,}\mathrm{m}$ range) for each method. This visual assessment reveals differences in spatial detail, forest boundary detection, and height estimation accuracy across the various approaches.

Panel of Figure 2 — Figure 2: Architecture of the Temporal-Swin-Unet. Skip connections are depicted as dashed arrows. The shape of the tensor in between a layer is shown in magenta: $C$ (channels), $T$ (timesteps), $0pt$ (height), $0pt$ (width), $Y$ (years), and $E$ (embedding dimension). In addition to the Video Swin Transformer Blocks ( Liu et al., 2022 ) , Encoder Layers have a Temporal Downsample (TD) layer at the end, and Decoder Layers a Temporal Skip Connection (TSC) at the beginning.

Method	MAE $\downarrow$	MSE $\downarrow$	RMSE $\downarrow$	MAPE $\downarrow$	$R^{2}$ $\uparrow$	$R_{\text{all}}^{2}$ $\uparrow$
Potapov et al. (2021)	1 9.11 [7.73]	185.52 [107.99]	13.62 [7.73]	54.90 [78.99]	0.50	0.70
Lang et al. (2023)	1 7.97 [6.67]	143.78 [ 1 81.04]	11.99 [6.67]	53.33 [76.84]	0.52	0.71
Pauls et al. (2024)	1 6.85 [6.41]	138.19 [ 1 60.75]	11.76 [6.41]	34.20 [33.16]	0.51	0.73
Tolan et al. (2024)	11.89 [9.09]	260.28 [182.25]	16.13 [9.09]	68.78 [57.80]	0.45	0.64
Ours	1 5.85 [4.90]	118.07 [ 1 36.16]	10.87 [4.90]	30.20 [33.32]	0.59	0.77

Table 1: Comparison of global-scale methods for 2020 regarding MAE (

\mathrm{m}

), MSE (

{\mathrm{m}}^{2}

), RMSE (

\mathrm{m}

), MAPE (%),

R^{2}

and

R_{\text{all}}^{2}

(on all labels, including labels below

5\text{\,}\mathrm{m}

). IQR is given in square brackets.

Figure 3: Left. Scatter plot showing the predicted height in 2018 against 2024. Disturbed pixels are identified by a decrease of more than $5\text{\,}\mathrm{m}$ between 2018 and 2024, marked red and excluded from the median aggregation. Right. Median height difference from 2018 to each year, binned in $1\text{\,}\mathrm{m}$ height classes. The right y-axis shows the height class distribution and area for these classes.

Figure 4: Examples of predicted tree height dynamics for two contrasting regions. Top : Le Landes (France) showing disturbance and regrowth patterns. Bottom : Amazonas (Brazil) with largely stable forest structure. Each block shows optical imagery (top row), predicted tree height (second row), and corresponding change maps from 2018 to 2024 (right column).

Figure 5: Left: Scattterplots showing the predicted height for 2020 vs GEDI labels with the correlation coefficient (

R^{2}

) and the correlation coefficient for labels exceeding

5\text{\,}\mathrm{m}

(

R^{2}_{5}

) indicated for each plot. Right: Histogram of the (predicted) values.

Figure 6: Error distribution analysis across height classes (

5\text{\,}\mathrm{m}

bins) for all baseline methods. Boxplots show mean absolute error for each height class, revealing that tall tree prediction remains challenging across all approaches, with errors increasing substantially for heights above

25\text{\,}\mathrm{m}

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Analysis of spatial autocorrelation between GEDI labels in our dataset. Correlation is given by the Pearson product-moment correlation coefficient in bins of

50\text{\,}\mathrm{m}

. The correlation starts to level off at around

300\text{\,}\mathrm{m}

400\text{\,}\mathrm{m}

Panel of Figure 8 — Figure 8: GEDI shot photon distribution. (a) The photon density of a single GEDI shot follows a normal distribution across both spatial dimensions. A Sentinel-2 pixel located directly at the center of the footprint receives 40.57% of the emitted laser energy. (b) Percentage of emitted photons received by sample trees at different positions. A tree with a $5\text{\,}\mathrm{m}$ radius centered directly under the GEDI footprint receives approximately 40% of the photons, while a tree of $7\text{\,}\mathrm{m}$ radius offset by $15\text{\,}\mathrm{m}$ from the center receives only 4.15%. Note that these values represent emitted photon distribution, not the reflected energy measured by the sensor. Actual measurements depend on material reflectance properties, ground return proportion, and the rh98 metric extraction, requiring substantially more than 2% photon coverage for counting a tree as the largest tree in a GEDI measurement.

Source	Origin	Res.	CRS	Resampling	Bands/Channels	Normalization
Sentinel-1	Google Earth Engine	10m	UTM	NN	VH (Asc/Des)	[-50, 1] in dB
Sentinel-2	AWS	10m	UTM	N/A	All w/o B10	[0, 1000]: 0 [0, 2000]: 1–4 [0, 6000]: 6–9 [0, 4000]: 5/10/11
ALOS Palsar-2	JAXA FTP	25m	EPSG:4326	Bilinear	HH/HV	[-50, 1] in dB
TandemX EDEM	DLR Geoservice	30m	EPSG:4326	Bilinear	—	[0, 7000]
TandemX FNF	DLR Geoservice	50m	EPSG:4326	NN	—	[0–2]
GEDI L2A	NASA Earthdata	$\sim$ 25m	EPSG:4326	N/A	rh98	—

Appendix

Table 2: Data sources used in ECHOSAT with their specifications and preprocessing parameters.

Figure 9: Sentinel-2 time series and temporal tree height prediction for five neighbouring pixels around a disturbance for the pretrained model (Left), our model (Middle) and model fine-tuned with a single linear regression model (Right). The simple linear regression model is fine-tuned on pseudo-labels produced by a linear interpolation of the pretrained model’s predictions.

Year	MAE $\downarrow$	MSE $\downarrow$	RMSE $\downarrow$	MAPE $\downarrow$	$R^{2}$ $\uparrow$	$R_{\text{all}}^{2}$ $\uparrow$
2019	6.09 [5.25]	123.44 [40.48]	6.09 [5.25]	0.30 [0.34]	0.59	0.77
2020	5.85 [4.90]	118.07 [36.16]	5.85 [4.90]	0.30 [0.33]	0.59	0.77
2021	5.41 [4.71]	96.03 [33.49]	5.41 [4.71]	0.30 [0.33]	0.63	0.79
2022	5.24 [4.75]	85.37 [34.34]	5.24 [4.75]	0.29 [0.31]	0.66	0.82
2023	5.56 [4.94]	102.86 [35.77]	5.56 [4.94]	0.29 [0.31]	0.62	0.79
2024	4.77 [4.26]	72.01 [27.78]	4.77 [4.26]	0.28 [0.28]	0.68	0.83

Appendix

Table 4: Year-wise comparison on error metrics regarding MAE (

\mathrm{m}

), MSE (

{\mathrm{m}}^{2}

), RMSE (

\mathrm{m}

), MAPE (%),

R^{2}

and

R_{\text{all}}^{2}

(on all labels, including labels below

5\text{\,}\mathrm{m}

). IQR is given in square brackets.

Figure 10: Error distribution analysis across height classes (

5\text{\,}\mathrm{m}

bins) for ECHOSAT predictions and different years. Boxplots show mean absolute error for each height class. The errors for smaller trees are very similar, with higher trees showing earlier years as being slightly more difficult than recent years.

Figure 11: Density scatterplot between ALS tree height labels and ECHOSAT predictions in the Le Landes forest plantation area in France. LiDAR campaigns are from the LiDAR HD of the French Geoservices and have been reprojected and downsampled to 10m and EPSG:32630 using max pooling. The prediction is taken from the corresponding year of the ALS campaign. As our model is only trained with GEDI labels and GEDI labels cannot differentiate between small trees, bushes and ground elevation, it sets the minimum height to $\approx$ $3\text{\,}\mathrm{m}$ . Bright colors indicate higher density.

Figure 12: Spatial distribution of errors for all labels (a) and different height bins (b-g). a) strongly correlates with known areas of taller trees (i.e. higher errors, see Figure 6 ), small trees (b-c) generally have low errors globally, whereas errors systematically growth with tree height (d-g). Interestingly, (1) the east coast of North America has lower errors compared to other areas and (2) the Amazonas rainforest and Congo Basin, known for tall trees, have smaller errors (g) compared to other regions.

Method	MAE $\downarrow$	MSE $\downarrow$	RMSE $\downarrow$	MAPE $\downarrow$	$R^{2}$ $\uparrow$	$R_{\text{all}}^{2}$ $\uparrow$
Potapov et al. (2021)	2.84 [0.56]	9.08 [3.08]	3.01 [0.56]	98.20 [ 1 0.00]	0.26	0.70
Lang et al. (2023)	3.04 [0.71]	13.30 [3.95]	3.65 [0.71]	104.08 [ 1 0.00]	0.28	0.71
Pauls et al. (2024)	1.08 [0.96]	2.99 [1.73]	1.73 [0.96]	37.04 [31.31]	0.49	0.73
Tolan et al. (2024)	2.87 [0.60]	8.93 [3.32]	2.99 [0.60]	99.04 [ 1 0.00]	0.16	0.64
Ours	0.51 [0.42]	1.82 [0.24]	1.35 [0.42]	15.99 [13.12]	0.39	0.77

Appendix

Table 5: Comparison of global-scale methods for 2020 only for labels below

5\text{\,}\mathrm{m}

regarding MAE (

\mathrm{m}

), MSE (

{\mathrm{m}}^{2}

), RMSE (

\mathrm{m}

), MAPE (%),

R^{2}

and

R_{\text{all}}^{2}

(on all labels, including labels above

5\text{\,}\mathrm{m}

). Please note that while GEDI measures ground elevation between

2\text{\,}\mathrm{m}

and

3\text{\,}\mathrm{m}

, Tolan et al. (2024) ; Lang et al. (2023) ; Potapov et al. (2021) either use other labels or manually set their ground predictions to

0\text{\,}\mathrm{m}

, thereby making a comparison here ineffective.

Parameter	Symbol	Value
		Pretraining	Finetuning
Number of years	$Y$	$7$
Number of timesteps	$T$	$84$ ( $=Y\cdot Months$ )
Number of input channels	$C$	$18$
Embedding dimension	$E$	$72$
Height / Width (in pixels)	$0pt$ / $0pt$	$96$

Appendix

Table 6: Training hyperparameters. Differences between pretraining and fine-tuning are highlighted.

Remote Sensing

Remote sensing uses satellite, aerial, and ground-based observations to measure and map Earth and other planetary surfaces.

See all 522 papers

CHMv2: Improvements in Global Canopy Height Mapping using DINOv3
Mar 6, 2026John Brandt, Seungeun Yi, Jamie Tolan +9Remote Sensing Remote Sensing Image Understanding
World Resources Institute, 10 G St NE #800, Washington, DC 20002, USA · Fundamental AI Research (FAIR), Meta, 75002 Paris, France · Meta, 1 Hacker Way, Menlo Park, CA 94025, USA +1
Uncertainty-aware tree height change regression
Jul 1, 2026Max Gaber, Dimitri Gominski, Jaime C. Revenga +3Uncertainty Quantification Geospatial Foundation Models
Department of Geosciences and Natural Resource Management, University of Copenhagen · Department of Mathematical Sciences, University of Copenhagen · Department of Computer Science and Electrical Engineering, University of Rostock +1
SERA-H: Super-Resolution of Sentinel Time Series for Fine-Scale Canopy Height Mapping
Dec 19, 2025Thomas Boudras, Martin Schwartz, Rasmus Fensholt +6Remote Sensing Image Super-Resolution Remote Sensing
Laboratoire des Sciences du Climat et de l’Environnement (LSCE), CEA, CNRS, UVSQ, Université Paris-Saclay, Gif-sur-Yvette, France · Department of Geography and Geology, University of Copenhagen, Øster Voldgade 10, Copenhagen, DK-1350, Denmark · Kayros SAS, Paris, 75009, France +2

Mar 6, 2026J/K move · Enter open · S save

CHMv2: Improvements in Global Canopy Height Mapping using DINOv3

John Brandt, Seungeun Yi, Jamie Tolan, Xinyuan Li+8

World Resources Institute, 10 G St NE #800, Washington, DC 20002, USA · Fundamental AI Research (FAIR), Meta, 75002 Paris, France · Meta, 1 Hacker Way, Menlo Park, CA 94025, USA+1

Remote Sensing Remote Sensing Image Understanding

Open paper AskPDF arXiv

Accurate canopy height information is essential for quantifying forest carbon, monitoring restoration and degradation, and assessing habitat structure, yet high-fidelity measurements from airborne laser scanning (ALS) remain unevenly available globally. Here we present CHMv2, a global, meter-resolution canopy height map derived from high-resolution optical satellite imagery using a depth-estimation model built on DINOv3 and trained against ALS canopy height models. Compared to existing products, CHMv2 substantially improves accuracy, reduces bias in tall forests, and better preserves fine-scale structure such as canopy edges and gaps. These gains are enabled by a large expansion of geographically diverse training data, automated data curation and registration, and a loss formulation and data sampling strategy tailored to canopy height distributions. We validate CHMv2 against independent ALS test sets and against tens of millions of GEDI and ICESat-2 observations, demonstrating consistent performance across major forest biomes.

Abstract

Remote Sensing

Explore similar work

CHMv2: Improvements in Global Canopy Height Mapping using DINOv3

Uncertainty-aware tree height change regression

SERA-H: Super-Resolution of Sentinel Time Series for Fine-Scale Canopy Height Mapping