A mask-based multimodal deep learning framework achieved the lowest test relative root mean square error in baseline comparisons when estimating tree diameter and allometry-derived biomass from drone imagery, according to a study conducted in South Korea. Evaluating Korean red pine across two experimental plots, the multimodal model recorded test relative root mean square errors of 10.93 percent and 11.09 percent for diameter at breast height across both leave-one-plot-out evaluation directions. For allometry-derived tree biomass, the baseline test relative root mean square errors reached 26.24 percent and 25.90 percent.
Accurate estimation of individual-tree biomass serves critical roles in forest inventory, carbon assessment, and fuel management. To develop the estimation workflow, researchers gathered field measurements alongside unmanned aerial vehicle RGB imagery from two Korean red pine plots located in South Korea. The architecture extracted crown metrics through Mask R-CNN delineation and merged those metrics with crown-specific image features inside a fully connected prediction head designed to estimate diameter at breast height and allometry-derived tree biomass.
The researchers evaluated the multimodal system against several alternative modeling techniques using a leave-one-plot-out validation design. Benchmarks included neural networks trained solely on images, neural networks trained solely on metrics, least squares regression, random forest, and XGBoost models. While the multimodal framework attained the lowest test relative root mean square error during baseline evaluations, paired comparisons did not consistently support lower multimodal error when matched against every individual benchmark model.
The investigation also revealed that model performance varied when researchers applied image resizing and tested alternative segmentation masks. These findings demonstrate support for integrating delineated crown metrics with image features for individual-tree parameter estimation under the specific forest conditions analyzed in the two plots. The authors emphasized that further evaluation remains necessary to verify how the framework performs across other forest conditions.
Subject of Research: Multimodal deep learning for individual-tree diameter and biomass estimation from UAV imagery
Article Title: Estimation of individual-tree DBH and allometry-derived biomass of Korean red pine using mask-based multimodal deep learning from UAV imagery
Article References: Kwon, K., Kim, S. Y., Lee, Y.-E., Cha, S., Kwon, C.-G., & Im, S.-K. (2026). Estimation of individual-tree DBH and allometry-derived biomass of Korean red pine using mask-based multimodal deep learning from UAV imagery. Scientific Reports. https://doi.org/10.1038/s41598-026-80383-7
Image Credits: AI Generated
DOI: 10.1038/s41598-026-80383-7
Keywords: Korean red pine, unmanned aerial vehicles, RGB imagery, deep learning, diameter at breast height, tree biomass, Mask R-CNN, crown metrics, forest inventory, carbon assessment, fuel management, leave-one-plot-out evaluation, random forest, XGBoost, least squares regression, multimodal framework, relative root mean square error, individual-tree estimation
Cite Scienmag News
Blake Davidson. (November 18, 2026). Multimodal Deep Learning Estimates Korean Red Pine Diameter and Biomass. Scienmag. https://scienmag.com/multimodal-deep-learning-estimates-korean-red-pine-diameter-and-biomass/
Blake Davidson. "Multimodal Deep Learning Estimates Korean Red Pine Diameter and Biomass." Scienmag, 18 November 2026, https://scienmag.com/multimodal-deep-learning-estimates-korean-red-pine-diameter-and-biomass/. Accessed 18 November 2026.
Blake Davidson. "Multimodal Deep Learning Estimates Korean Red Pine Diameter and Biomass." Scienmag. November 18, 2026. https://scienmag.com/multimodal-deep-learning-estimates-korean-red-pine-diameter-and-biomass/








