High Resolution Tree Height Mapping of the Amazon Forest using Planet NICFI Images and LiDAR-Informed U-Net Model

TL;DR

This study addresses the challenge of high-resolution canopy height mapping in the Amazon by training a LiDAR-informed U-Net regression model on Planet NICFI RGB-NIR imagery at $4.78$ m to predict per-pixel canopy height. The model is trained with LiDAR-derived canopy height models as references and validated across 3,436 patches, achieving a mean absolute error of $3.68$ m and handling heights up to $40$–$50$ m with minimal bias, outperforming existing global CHM products. The resulting Amazon canopy height map for 2020–2024 reveals a mean height of $22.09$ m, with tall forests concentrated in the Central Amazon and Guiana Shield, and demonstrates the potential to detect deforestation, logging, and regrowth through height changes. The approach enables high-resolution, large-scale monitoring of forest structure, informing carbon dynamics, biodiversity, and conservation planning, with future work aimed at improving cloud/shade handling and extending coverage to pantropical scales.

Abstract

Tree canopy height is one of the most important indicators of forest biomass, productivity, and ecosystem structure, but it is challenging to measure accurately from the ground and from space. Here, we used a U-Net model adapted for regression to map the mean tree canopy height in the Amazon forest from Planet NICFI images at ~4.78 m spatial resolution for the period 2020-2024. The U-Net model was trained using canopy height models computed from aerial LiDAR data as a reference, along with their corresponding Planet NICFI images. Predictions of tree heights on the validation sample exhibited a mean error of 3.68 m and showed relatively low systematic bias across the entire range of tree heights present in the Amazon forest. Our model successfully estimated canopy heights up to 40-50 m without much saturation, outperforming existing canopy height products from global models in this region. We determined that the Amazon forest has an average canopy height of ~22 m. Events such as logging or deforestation could be detected from changes in tree height, and encouraging results were obtained to monitor the height of regenerating forests. These findings demonstrate the potential for large-scale mapping and monitoring of tree height for old and regenerating Amazon forests using Planet NICFI imagery.

Figures (15)

• Figure 1: Location of the 3,060 LiDAR data points in the Amazon domain used in the validation of the Amazon forest tree canopy height model (shown in gold). As at least one image per flight was used in validation, the validation locations approximately represent the locations of all the LiDAR flights in the Amazon used in the study. The Amazon domain is partitioned into four regions according to Feldpausch2011.
• Figure 2: U-Net model architecture used for canopy height estimation from Planet NICFI images, adapted from Ronneberger2015. The number of channels is indicated above the cuboids, and the vertical numbers indicate the row and column sizes in pixels. The operations (convolutions, skip connections, max pooling, and upsampling) performed in each layer and their sizes are indicated by the colored arrows.
• Figure 3: Comparison of predicted versus observed height (m) for the 3,436 validation areas represented as density scatterplots for our canopy height model (a), Tolan's model (b), and Lang's model (c). Tolan's model and Lang's model, with native spatial resolutions of 0.5 m and 10 m respectively, were warped to our 4.78 m spatial resolution using the median and nearest neighbor algorithms, respectively. Each plot contains approximately 63 million points. The 1:1 line and the mean absolute error are depicted in white.
• Figure 4: Distribution of the observed height in the validation sample (a); distribution of the predicted height in the validation sample for our model, Tolan's model, and Lang's model (b); and distribution of the differences in predicted height in the validation sample between our model and Tolan's model, and between our model and Lang's model (c), respectively.
• Figure 5: Example of canopy height models observed in the validation dataset (column 1), predicted from our canopy height model (column 2), predicted from Tolan's model (column 3), and predicted from Lang's model (column 4). Tolan's model, Pauls's model and Lang's model, whose native spatial resolutions are 0.5 m, 10 m and 10 m, respectively, have been warped to our 4.78 m spatial resolution using the median and nearest neighbor algorithms, respectively.