PaperlayerView exact revision on arXiv

Source-linked AI summary

A high-resolution canopy height model of the Earth

Nico Lang, Walter Jetz, Konrad Schindler, Jan Dirk Wegner

arXiv:2204.08322v1cs.CVcs.LGeess.IV

TL;DR

Global canopy-height mapping requires both dense, high-resolution coverage and measurements of vertical structure, which no single source provides. The paper fuses GEDI references with Sentinel-2 imagery in a probabilistic deep learning framework to produce a global 10 m canopy-height map and uncertainty estimates. The resulting system reduces tall-canopy underestimation and supports worldwide conservation and monitoring applications.

  • Problem

    No single global data source combines dense high-resolution coverage with direct measurements of vertical vegetation structure for ecosystem and conservation analysis.

  • Method

    The authors fuse sparse GEDI LIDAR observations with Sentinel-2 optical images using probabilistic deep learning to retrieve canopy height and predictive uncertainty globally.

  • Results

    The method produces the first globally consistent, wall-to-wall canopy top height map of Earth at 10 m ground sampling distance and reduces underestimation bias for tall canopies.

  • Takeaways & Limitations

    The map supports biomass, biodiversity, climate, carbon, and forest-disturbance modelling and enables uncertainty-informed global monitoring.

  • Takeaways & Limitations

    Although mapped at 10 m ground sampling distance, effective feature resolution is lower because each pixel reflects the largest height within an approximately 25 m GEDI footprint and training supervision is sparse.

Abstract

from arXiv · show

The worldwide variation in vegetation height is fundamental to the global carbon cycle and central to the functioning of ecosystems and their biodiversity. Geospatially explicit and, ideally, highly resolved information is required to manage terrestrial ecosystems, mitigate climate change, and prevent biodiversity loss. Here, we present the first global, wall-to-wall canopy height map at 10 m ground sampling distance for the year 2020. No single data source meets these requirements: dedicated space missions like GEDI deliver sparse height data, with unprecedented coverage, whereas optical satellite images like Sentinel-2 offer dense observations globally, but cannot directly measure vertical structures. By fusing GEDI with Sentinel-2, we have developed a probabilistic deep learning model to retrieve canopy height from Sentinel-2 images anywhere on Earth, and to quantify the uncertainty in these estimates. The presented approach reduces the saturation effect commonly encountered when estimating canopy height from satellite images, allowing to resolve tall canopies with likely high carbon stocks. According to our map, only 5% of the global landmass is covered by trees taller than 30 m. Such data play an important role for conservation, e.g., we find that only 34% of these tall canopies are located within protected areas. Our model enables consistent, uncertainty-informed worldwide mapping and supports an ongoing monitoring to detect change and inform decision making. The approach can serve ongoing efforts in forest conservation, and has the potential to foster advances in climate, carbon, and biodiversity modelling.

1. Introduction

Global canopy height information is needed to support ecosystem, carbon, climate, and biodiversity assessment, but existing observations trade off coverage, resolution, and vertical-structure measurement. The study addresses this gap with a globally consistent 10 m canopy top height map from Sentinel-2 and GEDI.

  • Paper contribution: The paper presents a deep learning framework that maps canopy top height from publicly available optical satellite imagery across the entire global landmass.The approach is designed for consistent global deployment and extends beyond GEDI’s geographic coverage.
  • Paper contribution: The resulting 2020 product is the first global canopy top height map with 10 m ground sampling distance and is made publicly available for conservation and scientific modelling.The authors also release the underlying source code and models.
  • Why canopy height matters: Canopy height characterizes ecosystem structure and supports analysis of biomass, carbon stocks, habitat heterogeneity, biodiversity, and microclimate buffering.High-resolution canopy height models can inform climate-impact modelling and forest management.
  • Existing data limitations: GEDI provides globally distributed vertical-structure measurements but has limited geographic, spatial, and temporal coverage because it is a sparse sampling mission.Its samples restrict wall-to-wall products to a 1 km grid and cover at most 4% of the land surface.
  • Existing data limitations: Sentinel-2 and Landsat offer freely accessible, long-term global optical coverage at high spatial and temporal resolution, but do not directly measure vertical forest structure.Fusing optical imagery with GEDI observations can combine the strengths of both sources.

2. Deep learning approach

The approach scales deep learning to global canopy-height retrieval while targeting the underestimation of tall vegetation caused by optical-image saturation. It combines Sentinel-2 imagery, geographical information, reweighted training, and ensembles, with evaluation across held-out GEDI, biomes, and independent LVIS data.

  • Motivation: Existing canopy-height estimation tools tend to saturate around 25 to 30 m, underestimating tall canopies and impairing carbon-stock estimation in tall-forest regions.This limitation is particularly consequential in tropical forests, where tall trees have especially high biomass.
  • Model design: The model uses an ensemble of fully convolutional neural networks that transform Sentinel-2 optical images into dense 10 m canopy-height maps.The broader framework extends regional deep learning methods to global scale.
  • Model design: The CNN receives Sentinel-2 imagery and encoded geographical coordinates, learns from sparse GEDI supervision, and outputs canopy top height with predictive uncertainty.During map generation, predictions are produced for every input pixel despite training supervision being available only at valid GEDI-reference pixels.
  • Model design: The method addresses low-canopy sampling bias by reweighting losses inversely to sample frequency across 1-meter height intervals and aggregating repeated observations with CNN ensembles.The three proposed measures are reported to work best in combination.
  • Evaluation: 7.3 m aRMSE and -1.8 m aME were obtained with height balancing, while overall validation performance was 6.0 m RMSE and 1.3 m bias.The overall positive bias reflects slight overestimation of low canopies associated with improved performance on rare tall canopies.
  • Evaluation: Biome-level errors vary substantially, with average overestimation of approximately 2.5 m in mangroves, tundra, and tropical coniferous forests.Tundra estimates are affected by underrepresentation and noisy, outlier-contaminated GEDI reference data.
  • Evaluation: Independent evaluation across 12 LVIS areas produced an RMSE of 7.8 m and a mean error of 0.6 m.The RMSE was slightly higher but the bias lower than for held-out GEDI samples.

3. Modelling predictive uncertainty

The model quantifies predictive uncertainty alongside canopy-height estimates using an ensemble of deep CNNs and multi-date Sentinel-2 observations. Uncertainty-informed filtering can reduce error, but calibration varies across conditions.

  • Uncertainty estimation: Five independently initialized CNNs provide ensemble-based predictive uncertainty for canopy-height estimates.The ensemble captures uncertainty associated with model parameters and noisy inputs.
  • Uncertainty estimation: Ten satellite overpasses per year provide full coverage and redundant cloud-free observations for inference.Each observation is processed with a randomly selected ensemble CNN.
  • Uncertainty estimation: Filtering the 20% most uncertain estimates reduces overall RMSE by 13%, from 6.0 m to 5.2 m, while reducing bias by 23%, from 1.3 m to 1.0 m.This improvement comes at the cost of reduced completeness.
  • Uncertainty estimation: Inverse-variance weighting merges redundant predictions, but calibration deteriorates below 4 m standard deviation and varies by biome.Post-processing recalibration may therefore be advisable for particular applications and regions.
  • Uncertainty estimation: Predictive standard deviation is mapped alongside canopy top height in the 2020 global product.The uncertainty layer is visualized as predictive standard deviation at 10 m ground sampling distance.

4. Global canopy height mapping

The authors deploy the model across the 2020 Sentinel-2 archive to create a global 10 m canopy-height map with uncertainty information. The map resolves fine-scale and biome-level height distributions, while highlighting areas where predictions are more uncertain.

  • Global deployment: Producing the global map required approximately 160 terabytes of Sentinel-2 data and 27,000 GPU-hours, completed in 10 real-time days.The computation covered approximately 1.3×10^12 Sentinel-2-grid pixels.
  • Global deployment: The wall-to-wall product provides 10 m canopy-height and uncertainty information across the global landmass.Examples reveal forestry harvesting patterns, gallery forests, and dense tropical broadleaf forest structure.
  • Uncertainty patterns: Predictive uncertainty generally increases with estimated canopy height, but is also high in Alaska, northwestern Canada, and Tibet.The latter regions may contain conditions absent from training or challenging environments such as rare ecosystems and frequent cloud cover.
  • Global distributions: 5% of the global landmass is covered by vegetation taller than 30 m, compared with 41% above 5 m and 30% above 10 m.These estimates describe the 2020 global canopy-height distribution.
  • Global distributions: Protected areas tend to contain taller vegetation than the global average, yet only 34% of canopies above 30 m fall within protected areas.The comparison uses protected areas from the World Database on Protected Areas.
  • Biome distributions: Biome-level distributions differ substantially: tropical moist broadleaf forests are bimodal with a mode above 30 m, whereas mangroves are unimodal and reach above 40 m.The model also predicts reasonable canopy heights in tundra despite scarce and noisy reference data.

5. Discussion

The map supports biomass, biodiversity, and disturbance monitoring at global scale, while its effective spatial resolution and regional artifacts constrain interpretation of fine vegetation features.

  • Biomass and disturbance applications: Canopy height correlates strongly with aboveground carbon density in Sabah, supporting biomass and carbon-stock mapping.
  • Biomass and disturbance applications: Annual canopy-height differences correspond well with mapped wildfire extent in northern California and remain consistent in unaffected areas.
  • Biodiversity applications: Dense wall-to-wall coverage improves the use of sparse field samples by increasing spatial overlap and avoiding biases from low-resolution spatial averaging.
  • Resolution and map-quality constraints: Each pixel effectively reflects the largest canopy height within an approximately 25 m GEDI footprint, while sparse supervision and geolocation uncertainty reduce effective resolution.
  • Resolution and map-quality constraints: Tiling artifacts at northern high latitudes may degrade absolute height estimates, and their effects on differential quantities require further investigation.

6. Conclusion

The study delivers a globally consistent 10 m canopy-height map by combining Sentinel-2 imagery with sparse GEDI references. It reduces tall-canopy underestimation, provides uncertainty estimates, and supports updates and broader vegetation mapping.

  • The first globally consistent, wall-to-wall canopy top-height map of Earth uses Sentinel-2 imagery and has a 10 m ground sampling distance.
  • The model substantially reduces overall underestimation bias for tall canopies compared with previous global-scale canopy-height mapping.
  • Globally homogeneous predictive uncertainties and deployment without local calibration allow mapping both within and outside GEDI coverage.
  • Because trained models require only image data, canopy-height maps can be updated annually to track forest-goal progress.
  • Longer GEDI operation is expected to densify reference data and diminish predictive uncertainty.
  • The approach could be retrained or extended with additional missions to densely map biomass and other vegetation characteristics globally.
  • The work demonstrates that globally distributed reference data can enable deep learning to scale from regional remote sensing to global vegetation analysis.

7. Methods

The method fuses Sentinel-2 optical imagery with sparse GEDI canopy-height references in a probabilistic convolutional framework. It uses sparse pixel-wise supervision, ensembles, and repeated observations to estimate heights and uncertainty globally.

  • Data: Sentinel-2 supplies global optical imagery, while GEDI provides sparse 25 m-footprint canopy-top-height references between 51.6° north and south.
  • Data: Training combines GEDI and Sentinel-2 data by extracting 15×15-pixel, 150 m ×150 m image patches at GEDI footprint locations.
  • Evaluation: Independent airborne LVIS LIDAR canopy-top heights provide a second evaluation against data rastered to the Sentinel-2 10 m grid.
  • Model architecture: The model is a fully convolutional residual CNN using separable convolutions and spectral-textural features without downsampling.
  • Sparse supervision and training: Canopy-height retrieval is trained as pixel-wise regression with sparse supervision, minimizing loss only where GEDI reference values exist.
  • Uncertainty modelling: The network predicts a conditional Gaussian distribution per pixel, estimating mean and variance through a Gaussian negative-log-likelihood loss.
  • Uncertainty modelling: An ensemble of 5 CNNs and imagery from 10 acquisition dates captures model uncertainty and exploits repeated cloud-free observations.

7.5. Correction for imbalanced height distribution

The method addresses rare tall-canopy labels by reweighting samples during fine-tuning, while evaluating both prediction error and uncertainty calibration with balanced metrics.

  • Tall-canopy underestimation is partly attributed to the imbalanced distribution of reference labels, where large heights are rare.
  • Reference heights are binned into 1 m intervals, and sample influence is reweighted using moderated inverse-frequency weighting.The empirically preferred weighting uses the square root of inverse frequency and fine-tunes only the final mean-height regression layer.
  • Prediction performance is measured with RMSE, MAE, and ME, with negative ME indicating systematic underestimation.
  • Balanced aRMSE, aMAE, and aME average errors across 5 m height intervals rather than weighting intervals by their sample counts.
  • Uncertainty calibration compares binned predictive uncertainty with empirical error, using AUCE to give each bin equal weight.For calibration, empirical error is represented by RMSE and bin uncertainty by predictive standard deviation derived from RMV.

7.7. Global map computation

Global canopy height mapping uses Sentinel-2 imagery organized into worldwide tiles, followed by minimal masking and standardized tile-based product release. The computation is performed on a GPU cluster, with energy use contextualized through alternative cloud scenarios.

  • Sentinel-2 imagery is processed in 100 km×100 km tiles, with 18,011 tiles covering global landmass excluding Antarctica.Relevant satellite orbits are selected per tile according to the smallest number of empty pixels.
  • Built-up areas, snow, ice, and permanent water bodies are masked as no data using ESA World Cover classification.The product is released as 3°×3° geographic tiles to simplify integration with existing workflows.
  • The global map is computed on a Swiss GPU cluster, with carbon emissions estimated for AWS scenarios in Stockholm and Ohio.

Data and code availability

The 2020 global canopy height map, interactive viewer, source code, and trained models are made available for download and reuse.

  • The global canopy height map is available for download and in Google Earth Engine, alongside source code and trained models released through the project page.An interactive browser application provides map exploration.

Appendix A.1. Comparison to UMD canopy height map

The map is compared with the UMD canopy height product using independent LVIS airborne-LIDAR reference data, while additional appendix figures examine geographic error, height-specific performance, uncertainty calibration, and protected-area patterns.

  • Comparison with UMD canopy height map: The UMD comparison uses independently created LVIS canopy top height data and restricts evaluation to LVIS regions within GEDI coverage.This restriction reflects the UMD map’s absence north of 51.6° latitude.
  • Comparison with UMD canopy height map: The map is more accurate than UMD across all reported error metrics, with the largest difference in mean error.
  • Comparison with UMD canopy height map: UMD shows −4.8 m underestimation, whereas the presented map shows 0.2 m overestimation bias.
  • Appendix error analyses: Geographical error analysis reports RMSE, MAE, and ME at 1° resolution on held-out GEDI validation data.Negative ME denotes underestimation bias.
  • Appendix error analyses: Independent LVIS evaluation compares Sentinel-2 predictions with reference canopy top height across all LVIS areas, GEDI-covered regions, and regions north of GEDI.
  • Uncertainty and spatial analyses: Calibration plots relate estimated predictive uncertainty to empirical error, while uncertainty filtering evaluates changes in overall error metrics.
  • Uncertainty and spatial analyses: Biome-level confusion plots and protected-area analysis extend the appendix evaluation to ecological and conservation contexts.
Loading 2204.08322v1…