Traditional methods for measuring forests rely on labor-intensive ground surveys, which are costly and limited in scale. While drones can capture rich, ultra-high-resolution images, single-modal data like standard red, green, blue (RGB) photographs often struggle to tell species apart when they look similar or grow closely together, especially in complex natural environments. Light detection and ranging (LiDAR) technology, which provides detailed three-dimensional (3D) structural information, offers a solution but is difficult to integrate with two-dimensional (2D) imagery. Based on these challenges, there is a need for a sophisticated, unified approach that can seamlessly combine these different streams of information to identify trees and measure their vital statistics automatically and accurately.
A team of researchers from the Institute of Forest Resource Information Techniques at the Chinese Academy of Forestry, along with collaborators at the Qilian Mountain National Park, has developed an advanced AI model to address this. Their findings, published (DOI: 10.1016/j.plaph.2026.100171) on September 7, 2026, in the journal Plant Phenomics, detail a novel framework called SAMFormer that leverages multimodal drone data for instance segmentation of individual trees. The technology not only identifies species with high accuracy but also enables the large-scale extraction of structural parameters like height, diameter at breast height (DBH), and biomass, which are essential for calculating carbon stock.
The core innovation is SAMFormer (Self-Attention-guided Spectral-Structural Multimodal Fusion Transformer), which functions like a highly intelligent digital forester. It uses a dual-stream network to independently process RGB images and LiDAR-derived features, such as the Canopy Height Model (CHM). The model's brilliance lies in two key modules: an Adaptive Feature Enhancement Module (AFEM) that uses spatial and channel attention to sharpen the focus on canopy features while suppressing background noise, and a Cross-Modal Fusion Module (CMFM) that acts as a translator, allowing the two different data types to “talk” to each other and learn complementary features. The results are impressive: the model achieved an F1-score of 86.3% and mAP@0.5 of 88.0%, significantly outperforming single-modal inputs and mainstream models like Mask R-CNN. This high precision allowed the team to create species-specific maps of structural parameters and carbon stock, revealing crucial ecological insights: competition among trees showed a significant negative correlation with their carbon stock. Trees under intense competition adapt by reducing their growth, limiting canopy expansion, and ultimately storing less carbon. The study also found that certain mixed-species forests—like those combining conifers and broadleaf trees—stored more carbon than monocultures, while mixing inappropriate species could have the opposite effect.
“We’ve essentially given forest managers a new pair of eyes that can see both the leaves and the wood,” the authors said. “For the first time, we can map not just the location of a tree, but its exact species, its structural details, and its contribution to carbon storage, all from the air. This high-throughput, non-destructive method is a game-changer for forest phenotyping. It offers a practical way to move beyond stand-level averages and truly understand what's happening in our forests at the scale of an individual tree, which is where the real ecological action is.”
This framework has immediate and far-reaching applications for "climate-adaptive" forest management. The detailed, wall-to-wall maps of species-specific structural parameters enable managers to pinpoint areas of high competition and optimize harvest planning to reduce pressure on trees, thereby improving overall forest health and resilience. The ability to distinguish between species and accurately quantify their biomass allows for precise carbon accounting, which is essential for carbon credit verification and monitoring. This technology provides a scalable, objective, and efficient tool to support global efforts in conservation and mitigation, transforming how we inventory, manage, and protect our planet's vital forest ecosystems.
###
References
DOI
10.1016/j.plaphe.2026.100171
Original Source URL
https://doi.org/10.1016/j.plaphe.2026.100171
Funding information
This work was funded by the National Natural Science Foundation of China (32271877) and the National Key Research and Development Program of China (2022YFE0128100).
About Plant Phenomics
Plant Phenomics publishes breakthrough research that integrates genomics, genetics, physiology, molecular biology, bioinformatics, statistics, mathematics and computer science to advance plant phenotyping—from the cellular level to entire populations. The journal tackles key scientific challenges in phenomics by showcasing cutting‑edge technologies for data acquisition, management, interpretation and modelling. Its scope spans practical applications across all plant sciences, aiming to translate novel phenotyping tools into real‑world solutions for breeding, crop management and fundamental plant biology. By fostering interdisciplinary innovation, Plant Phenomics serves as a hub for researchers developing high‑throughput, accurate and scalable methods to decipher complex plant traits, ultimately accelerating progress in agriculture and plant science.