Source-linked AI summary
Country-wide high-resolution vegetation height mapping with Sentinel-2
Nico Lang, Konrad Schindler, Jan Dirk Wegner
TL;DR
The paper addresses the difficulty of mapping vegetation height over large areas at high spatial resolution. It uses Sentinel-2 multispectral imagery and deep CNN regression to produce country-scale maps, demonstrating 10 m mapping with low error across tropical and European settings.
Problem
Direct tree-height measurement does not scale to large areas and high spatial resolution, while airborne LiDAR is costly and limited in geographic coverage.
Method
A deep CNN regresses vegetation height from Sentinel-2 multispectral reflectance, learning spectral and neighborhood texture features at 10 m resolution.
Results
The approach demonstrates country-scale vegetation-height mapping at 10 m resolution, retrieves heights up to approximately 55 m, and achieves low error in both tropical and central European settings.
Takeaways & Limitations
The method supports high-resolution vegetation-structure applications, frequent updates, and potentially reduced reliance on expensive LiDAR flight campaigns.
Takeaways & Limitations
In Gabon, existing high-resolution canopy-height maps with a 15 m cutoff largely degenerate into binary forest layers because most forest exceeds 30 m.
Abstract
from arXiv · showhide
Sentinel-2 multi-spectral images collected over periods of several months were used to estimate vegetation height for Gabon and Switzerland. A deep convolutional neural network (CNN) was trained to extract suitable spectral and textural features from reflectance images and to regress per-pixel vegetation height. In Gabon, reference heights for training and validation were derived from airborne LiDAR measurements. In Switzerland, reference heights were taken from an existing canopy height model derived via photogrammetric surface reconstruction. The resulting maps have a mean absolute error (MAE) of 1.7 m in Switzerland and 4.3 m in Gabon (a root mean square error (RMSE) of 3.4 m and 5.6 m, respectively), and correctly estimate vegetation heights up to >50 m. They also show good qualitative agreement with existing vegetation height maps. Our work demonstrates that, given a moderate amount of reference data (i.e., 2000 km$^2$ in Gabon and $\approx$5800 km$^2$ in Switzerland), high-resolution vegetation height maps with 10 m ground sampling distance (GSD) can be derived at country scale from Sentinel-2 imagery.
1. Introduction
The paper addresses the difficulty of obtaining dense, high-resolution vegetation-height maps by using Sentinel-2 imagery and deep CNN regression. It demonstrates country-wide canopy-height mapping at 10 m resolution and extends sensitivity to tropical forests up to approximately 55 m.
- Direct tree-height measurements do not scale to large areas and high spatial resolution, while airborne and spaceborne LiDAR have coverage or sampling limitations.In-situ measurements are limited to sample plots and logging sites; airborne LiDAR is costly and regional, whereas spaceborne LiDAR is sparse in space and time.
- Existing regression from multispectral satellite images has produced tree-height maps at resolutions down to 30 m using Landsat and reference heights.
- 10 m country-wide canopy-height mapping is demonstrated by regressing Sentinel-2 multispectral data with a deep convolutional neural network.At this resolution, neighborhood patterns related to shadowing, roughness, and species distribution provide texture information beyond an individual pixel’s spectral signature.
- Texture patterns are particularly important in high tropical forest and extend the regressor’s sensitivity to vegetation heights of approximately 55 m.
- The CNN regresses canopy height for Gabon and Switzerland from 13-channel Sentinel-2 bottom-of-atmosphere reflectance, using airborne LiDAR and photogrammetric stereo-matching references.
- The work claims the first large-scale vegetation-height mapping from optical satellites at 10 m GSD, with retrieval beyond the saturation level of existing high-resolution maps.The authors also report that they are unaware of other optical-satellite canopy-height work using deep CNNs.
2. Related work
Prior vegetation-height mapping has relied on LiDAR, optical or radar satellite regression, and reference data, but existing approaches face coverage, sampling, or height-saturation limits. This paper’s related-work context motivates using spatial texture with high-resolution Sentinel-2 imagery.
- Remote sensing of vegetation height: Airborne LiDAR directly measures canopy height accurately but is time-consuming and expensive for wide-area coverage.
- Remote sensing of vegetation height: Spaceborne LiDAR scales coverage at the cost of sparse spatial sampling, as illustrated by GLAS profiles with a 70 m footprint.GEDI is described as a newer system with 25 m footprints and denser along-track and across-track spacing than GLAS.
- Remote sensing of vegetation height: Dense canopy-height maps can be produced by regressing optical satellite imagery against tree heights derived from LiDAR or sufficiently numerous in-situ observations.
- Remote sensing of vegetation height: Optical pixel signatures provide proxies such as shadowing, vegetation type, and density, supporting canopy-height inference from monocular imagery.
- Remote sensing of vegetation height: Earlier optical approaches often used single-pixel spectral data, leaving planimetric texture as a potential additional proxy for vegetation structure at high resolution.
- Remote sensing of vegetation height: Radar-based methods achieve approximately 2 m RMSE for trees up to 30 m, but a tropical PALSAR-1 map resolves heights only up to 15 m.
- Sentinel-2 and deep learning: Sentinel-2 had been used relatively little for forest properties, with one cited tree-height study training a two-layer perceptron on fewer than 200 field plots.
- Sentinel-2 and deep learning: Deep CNNs had become dominant in image analysis and remote sensing, though remote-sensing applications were predominantly classification rather than regression.
3. Data
The study uses Sentinel-2 multispectral imagery and region-specific reference data from Gabon and Switzerland. The Swiss canopy-height reference is photogrammetric and includes slope-related errors that require filtering.
- Sentinel-2: Sentinel-2 comprises two satellites with a combined 5-day revisit time and 13 spectral bands at 10 m, 20 m, and 60 m resolutions.Four bands provide 10 m GSD in blue, green, red, and near-infrared wavelengths.
- Sentinel-2: The study queried Sentinel-2 Level 1C tiles with at most 70% cloud cover, using imagery temporally close to Gabon reference acquisitions and Switzerland’s 2016 leaf-on season.
- Gabon / tropical Africa: NASA’s LVIS campaign collected full-waveform airborne LiDAR scans in five Gabon regions, including tropical and mangrove forests with canopy heights exceeding 71.1 m and 85 m.The LiDAR beams had an 18 m footprint and approximately 10 m along-track and across-track spacing.
- Switzerland / Alps: The Swiss reference canopy-height model was derived by subtracting a digital terrain model from a photogrammetric digital surface model and masking buildings.The original model had 1 m GSD and was reprojected and resampled to 10 m GSD.
- Switzerland / Alps: The Swiss canopy-height model has a terrestrial-validation RMSE between 3.6 m and 5.0 m after filtering outliers.
- Switzerland / Alps: Samples above 40 m were excluded from Swiss testing because they occurred on steep slopes and were considered unrealistic for the observed alpine forests.The excluded samples represented 0.03% and 0.08% of the two Swiss regions, while lower-height mis-registration biases were ignored.
4. Method
The method uses a deep, fully convolutional CNN to regress canopy height from atmospherically corrected and normalized Sentinel-2 imagery, learning spectral, spatial-context, and texture features. Training uses geographically separated validation and test regions, while inference tiles large images and masks cloud, water, and— in Switzerland—snow pixels.
- Preprocessing: 13-channel Sentinel-2 bottom-of-atmosphere reflectance is normalized per channel before CNN training.Atmospheric correction homogenizes image values across sensing dates and regions; channel standardization gives features comparable magnitudes.
- Deep regression network: The network adapts Xception with an entry block and 18 residual depthwise-separable convolution blocks.The entry block increases channel depth to 728, while separable convolutions learn spectral, spatial-context, and texture features.
- Objective and resolution: The final pointwise convolution produces one canopy-height value per pixel by minimizing mean square error with optional L2 weight decay.The architecture has 19,604,225 trainable weights and retains full spatial resolution because convolutions use stride 1 without pooling.
- Training: Training samples are 15 × 15-pixel patches, with missing ground-truth pixels excluded from each patch’s loss.Optimization uses mini-batch stochastic gradient descent on randomly sampled patches.
- Training and validation: Geographically separated test regions remain unseen during training, while validation loss monitors generalization and possible overfitting.Training uses about 50,000 iterations for one region and about 250,000 iterations across all regions, with validation loss evaluated every 500 iterations.
- Inference and evaluation: Large-scale inference recomposes 128×128-pixel tiles with 8-pixel overlap and masks pixels with cloud probability above 10%.Water pixels are excluded from quantitative evaluation, and snow pixels are additionally excluded in Switzerland.
5. Results and discussion
The CNN was evaluated across seven regions, with country-level accuracy, cross-region generalisation, spectral-band ablations, and texture-feature tests. Predictions generally matched spatial vegetation patterns, but smoothing, cloud effects, high-height underestimation, and reference-data outliers limited performance.
- Results for reference areas: Gabon overestimated heights around 30 m and underestimated vegetation between 50–70 m, while Switzerland showed an approximately −2.5 m offset above 15 m.In Gabon, rainforest reflectance and texture appeared to saturate at high canopy heights; Swiss reference data also contained implausible steep-slope outliers.
- Results for reference areas: The maps recovered broad spatial distributions and local variations, including canopy structures smaller than 100 m, but fine structures were over-smoothed.Very high trees above 50 m were underestimated, while nearby lower vegetation tended to be overestimated.
- Generalisation across time and space: Cross-region testing in Switzerland reduced performance by only 0.4 m and 0.2 m, with the drop probably caused mainly by roughly half as much training data.The cross-validation trained on one region and tested on the other, without nearby test-area data during training.
- Ablation study: spectral bands: All alternative band selections performed significantly worse in Gabon, whereas Switzerland slightly improved with RGBN and remained effective with visible RGB alone.The authors caution that the Swiss improvement cannot be assessed statistically from one test region, while near-infrared alone was insufficient.
- Influence of texture features: Removing spatial context increased MAE from 3.7 m to 6.0 m in GA3 and from 2.6 m to 3.6 m in CH2, especially at 40–60 m heights.The restricted 1×1-kernel CNN retained the same architecture apart from its inability to use texture or spatial context.
6. Conclusion
The study demonstrates country-scale vegetation height mapping at 10 m resolution from Sentinel-2 imagery, while identifying limits in temporal information, geographic generalisation, and training-data coverage.
- 6. Conclusion: 10 m GSD maps can be derived from Sentinel-2 imagery for country-scale vegetation height mapping.The approach uses a data-driven CNN that exploits spatial context and texture features.
- 6. Conclusion: Single cloud-free Sentinel-2 images can produce vegetation height maps with good accuracy, without requiring long cloud-free time series.The paper also notes that Sentinel-2 provides a new image every 5 days.
- 6. Conclusion: The authors identify targeted forest protection and fine-grained biodiversity studies as applications enabled by localized vegetation-structure information.They also suggest Sentinel-2 retrieval could sometimes reduce the need for expensive LiDAR flight campaigns.
- 6. Conclusion: Rudimentary use of multi-temporal information is a limitation because spectral time series could potentially improve canopy-height predictions.The current approach does not rely on long cloud-free time series, yet achieves satisfactory performance.
- 6. Conclusion: Global extension remains constrained by available training data and limited generalisation to unseen world regions.The authors propose diverse global training data and consider GEDI as one possible source, despite its 25 m footprint.