- Tree Level
Amazon Basin canopy height
Access high-resolution data on canopy height for the entire Amazon forest
Lead scientist
This dataset is available as an open resource via the Open Data Registry on AWS, and via the Google Earth Engine Community Catalog.

Insights on the Amazon tree canopy
This dataset provides high-resolution (~4.78 m) tree canopy height for the Amazon forest from 2020-2024. The data was generated using a U-Net deep learning model adapted for regression, trained with Planet NICFI satellite images and informed by LiDAR-derived canopy height models (CHMs).
The resulting map reveals substantial spatial variation in Amazon forest structure. Across forested pixels, the estimated mean canopy height was 22.09 meters, the median was 22.25 meters, and the 97.5th percentile was 32.10 meters. Taller forests were concentrated in parts of the Central Amazon and Guiana Shield, while shorter canopies were associated with wetlands, major rivers, savannas, rocky areas, high elevations, roads, and deforested landscapes.
The publicly available product represents mean canopy height for the 2020–2024 period. Year-specific canopy height products for selected parcels or areas of interest can be delivered on demand.
A canopy height model trained on airborne LiDAR
Input data
The canopy height model uses normalized analytic Planet NICFI imagery basemaps provided through Norway’s International Climate and Forest Initiative and Planet. The imagery includes four spectral bands: Red; Green; Blue; and Near-infrared.
The native spatial resolution is approximately 4.78 meters. Planet NICFI scenes are distributed as approximately 20 × 20-kilometer tiles. Each monthly or biannual basemap is assembled from selected daily observations to reduce cloud contamination.
The study processed 56 available NICFI dates. Before model training, the spectral bands were converted to an unsigned 8-bit RGB-NIR composite. No additional atmospheric correction was performed because the NICFI normalized basemaps do not guarantee absolute radiometric consistency among dates and sensors.
LiDAR-derived canopy height models
Airborne LiDAR data used to train and validate the model were obtained from: the Brazilian Agricultural Research Corporation, or Embrapa; the Sustainable Landscapes Brazil program; the Biomass Estimation of the Amazon project coordinated by INPE and partner institutions; LiDAR surveys in the Brazilian Amazon and one study area in Peru; and the São Paulo Metropolitan Region LiDAR dataset.
The original LiDAR point clouds were processed into one-meter canopy height models. Digital terrain and surface models were generated from the classified point clouds, and canopy height was calculated as the difference between the digital surface model and digital terrain model.
Google Open Buildings footprints for South America were used to remove buildings and other constructed features that could otherwise be incorrectly represented as vegetation height in the LiDAR training data. Building polygons were buffered by approximately 5 meters outside cities and 15 meters inside cities before being rasterized and removed from the canopy height models.
Canopy-height regression
A classical U-Net convolutional neural network was adapted to predict continuous canopy height values rather than discrete land-cover classes.
Key model characteristics included:
- approximately 35 million trainable parameters;
- four-band RGB-NIR inputs;
- input patches of 256 × 256 pixels;
- 4.78-meter input and output resolution;
- a single-band continuous canopy height output;
- TensorFlow 2.10 and Keras implementation through R;
- Mean Squared Error as the loss function;
- Mean Absolute Error as the evaluation metric;
- Adam optimizer with a learning rate of 0.0001;
- batch size of 32; and
- training for as many as 1,000 epochs.
The loss function was weighted using a binary mask identifying pixels with valid LiDAR observations. This allowed partially covered LiDAR flight areas to be included without treating missing reference pixels as zero-height vegetation. No artificial data augmentation was used.
Training and validation samples
The final modeling dataset contained 50,724 image patches. The validation dataset represented approximately 7.2% of the complete sample. It included 3,060 patches distributed across the Amazon forest domain and 376 patches from the Brazilian Atlantic Forest domain. To improve spatial independence and representation, the validation sample was constructed by selecting approximately one patch from each Planet NICFI tile overlapping the available LiDAR datasets.
Large-area prediction
Predictions from multiple cloud-free observations were combined to produce the final mean canopy height composite. Clouds, cloud shadows, haze, deforested areas predating the target period, and areas undergoing recent regeneration were filtered when constructing the stable multiyear canopy height product. The number of valid observations was retained during processing as an indicator of confidence and data availability.
Validation approach and results
Validation was performed against independent LiDAR-derived canopy height models across 3,436 spatially distributed validation areas. Comparisons included approximately 63 million pixel observations after harmonizing the reference and predicted datasets to the same 4.78-meter grid.
The model demonstrated relatively low systematic bias across most of the canopy height distribution (Mean Absolute Error: 3.68 m and Correlation for area-level median canopy height: Pearson r = 0.927). It slightly overestimated some canopy heights below approximately 25 meters and increasingly underestimated the tallest canopy percentiles. Trees taller than 50 meters were generally underestimated, although these represented only about 0.06% of validation pixels.
At the validation-area level, median predicted canopy height closely followed the LiDAR observations. The 95th-percentile height showed increasing underestimation for the tallest forests, reaching differences of approximately 5–10 meters in some areas.
The model outperformed the global canopy height products evaluated by the authors on the same validation dataset. However, performance may be reduced in areas with persistent clouds, strong terrain shadows, image blur, or limited numbers of valid optical observations, particularly along parts of the Andes and western Amazon.
Data to analyze forest structure and detect forest activity
Canopy height data allows for single-year analysis of biomass, carbon, forest structure, and ecosystem; and annual time series enable monitoring of deforestation, selective logging, and regeneration.
Forest biomass and carbon assessment
Canopy height can support aboveground biomass and carbon-stock estimation when integrated with canopy cover, wood-density information, forest type, field inventories, and locally calibrated allometric models. The dataset can also help characterize variation in forest structure that is not captured by coarser regional or global canopy height products.
Forest structure and ecosystem analysis
The product can be used to:
- Quantify regional variation in canopy height
- Identify areas containing exceptionally tall forests
- Characterize canopy-height distributions
- Evaluate forest structural complexity
- Assess habitat structure
- Analyze relationships among canopy height, soils, topography, hydrology, climate, and disturbance
Monitoring deforestation
Abrupt decreases in predicted canopy height can help identify clear-cutting and forest conversion. Height information can complement conventional tree-cover change products by providing an estimate of the vertical structure removed during a disturbance.
Detecting selective logging and degradation
Monthly Planet NICFI predictions showed the potential to locate canopy-height reductions associated with selective logging. The paper demonstrated that negative month-to-month height differences corresponded spatially with newly logged areas, although illumination, geolocation, phenology, and cloud artifacts can introduce noise.
Monitoring regenerating forests
Time series of predicted canopy height can be used to monitor secondary forest recovery following deforestation or agricultural abandonment. The study demonstrated examples in which regrowing vegetation increased from low initial heights to approximately 15–20 meters over several years. Further validation is required before interpreting short-term predicted differences as precise measurements of annual tree growth.
References
Wagner, Fabien H., et al. "High resolution tree height mapping of the amazon forest using planet nicfi images and lidar-informed u-net model." arXiv preprint arXiv:2501.10600 (2025).
Wagner, F.H., Dalagnol, R., Carter, G., Hirye, M.C.M., Gill, S., Takougoum, L.B.S., Favrichon, S., Keller, M., Ometto, J.P.H.B., Alves, L., Creze, C., George-Chacon, S.P., Li, S., Liu, Z., Mullissa, A., Yang, Y., Santos, E.G., Worden, S.R., Brandt, M., Ciais, P., Hagen, S.C. and Saatchi, S. (2026), Wall-to-wall Amazon forest height mapping with Planet NICFI, Aerial LiDAR, and a U-Net regression model. Remote Sens Ecol Conserv, 12: 323-348. https://doi.org/10.1002/rse2.70041 (opens in a new tab)
Wagner, F.H.; Breunig, F.M.; Balbinot, R.; Silva, E.A.; Soares, M.C.; Kramm, M.A.; Hirye, M.C.M.; Carter, G.; Dalagnol, R.; Hagen, S.C.; et al. Monitoring the Early Growth of Pinus and Eucalyptus Plantations Using a Planet NICFI-Based Canopy Height Model: A Case Study in Riqueza, Brazil. Remote Sens. 2025, 17, 2718. https://doi.org/10.3390/rs17152718 (opens in a new tab)
Data access
Team
Page last updated Nov 16, 2026