Li et al. (2026) Deep learning tree and forest biomass from sub-meter resolution optical imagery

Identification

• Journal: Remote Sensing of Environment
• Year: 2026
• Date: 2026-07-30
• Authors: Sizhuo Li, Martin Brandt, Xiaoye Tong, Stefan Oehmcke, Christian Igel, Florian Reiner, Fabian Gieseke, Thomas Nord‐Larsen, Rasmus Fensholt, Jerome Chave, Philippe Ciais
• DOI: 10.1016/j.rse.2026.115539

Research Groups

• Department of Geosciences and Natural Resource Management, University of Copenhagen, Denmark
• Department of Computer Science, University of Copenhagen, Denmark
• Department of Information Systems, University of Münster, Germany
• Laboratoire Evolution et Diversité Biologique, CNRS, UPS, IRD, Université Paul Sabatier, France
• Laboratoire des Sciences du Climat et de l'Environnement, CEA, CNRS, UVSQ, Université Paris-Saclay, France

Short Summary

The study evaluates the ability of Convolutional Neural Networks (CNNs) to directly estimate forest above-ground biomass (AGB) from sub-meter resolution RGB optical imagery. The results demonstrate that CNNs can achieve high predictive performance ($R^2 = 0.71$) without requiring LiDAR-derived tree height data, approaching the accuracy of traditional height-based models.

Objective

• To determine if CNNs can interpret spatial semantic patterns in high-resolution optical RGB images to directly estimate forest AGB, thereby reducing the dependency on LiDAR data for structural measurements.

Study Configuration

• Spatial Scale: Sub-meter resolution (down to single-pixel level) and forest inventory plots.
• Temporal Scale: Not explicitly specified, although the study notes that optical imagery typically offers higher temporal frequency than LiDAR data.

Methodology and Data

• Models used: Convolutional Neural Networks (CNN), Random Forest (benchmark), and a learnable allometric model based on smooth min-max networks.
• Data sources: Sub-meter resolution optical RGB imagery, forest inventory plots (field measurements), and LiDAR data (used for benchmark comparisons).

Main Results

• The CNN model achieved a high performance of $R^2 = 0.71$ using only optical RGB imagery.
• The CNN performance approached that of a Random Forest model that had access to tree height information ($R^2 = 0.8$).
• The CNN demonstrated the ability to implicitly interpret biomass composition at the tree level rather than relying on aggregated parameters.

Contributions

• Provides a method for direct AGB estimation from high-resolution RGB imagery, bypassing the need for LiDAR.
• Introduces a learnable allometric model using smooth min-max networks to flexibly adjust relationships between field biomass and remotely sensed features.
• Enhances spatial and temporal flexibility in forest biomass monitoring by leveraging the higher availability of optical imagery over LiDAR.

Funding

• Not specified in the provided text.

Citation

@article{Li2026Deep,
  author = {Li, Sizhuo and Brandt, Martin and Tong, Xiaoye and Oehmcke, Stefan and Igel, Christian and Reiner, Florian and Gieseke, Fabian and Nord‐Larsen, Thomas and Fensholt, Rasmus and Chave, Jerome and Ciais, Philippe},
  title = {Deep learning tree and forest biomass from sub-meter resolution optical imagery},
  journal = {Remote Sensing of Environment},
  year = {2026},
  doi = {10.1016/j.rse.2026.115539},
  url = {https://doi.org/10.1016/j.rse.2026.115539}
}