KNOXVILLE TN. December 26 2025 /24-7PressRelease/ – Forests and plantations are vital in the carbon sequestration process, but accurately tracking their growth is costly and time-consuming. Researchers developed an artificial intelligence model (AI) that can produce high-resolution maps of canopy height using standard RGB images. This method, which combines a large-vision foundation model and self-supervised improvement, achieves near-lidar precision, making it possible to monitor forest biomass, carbon storage, over large regions, with low cost.
It is crucial to monitor forest canopy structure in order to understand global carbon cycles and manage plantation resources. Lidar systems are expensive and complex, but they provide high-quality height measurements. Optical remote sensing is often unable to deliver the precision needed for plantations of small scale. While deep learning has improved the accuracy of canopy estimation, it still requires massive datasets that are labeled and can often result in a loss of fine spatial detail. Global models also struggle to adjust to plantation landscapes that are fragmented and have uniform tree structure. A cost-effective and high-resolution approach to mapping canopy heights and biomass is urgently needed due to these problems.
The team, which includes researchers from Beijing Forestry University and Manchester Metropolitan University as well as Tsinghua University, has created a vision model driven by artificial intelligence that can estimate tree heights with sub-meter precision from RGB satellite imagery. Published (DOI: 10.34133/remotesensing.0880) in the Journal of Remote Sensing on October 20, 2025, the study introduces a novel framework that combines large vision foundation models (LVFMs) with self-supervised learning. This approach solves the problem of cost, precision and scalability when it comes to forest monitoring. It is a powerful tool that can be used for tracking carbon absorption under programs such as China’s Certified Emission Reduction Program.
Researchers created a network of modules for canopy height estimation. These included a feature extraction unit powered by DINOv2’s large vision model, an enhanced feature module that retains fine spatial detail, and lightweight convolutional height estimator. Comparing the model to airborne lidar measurements it achieved an absolute error mean of just 0.09m, and an R2 value of 0.78. This is a significant improvement over traditional CNN or transformer-based methods. The model also achieved a 90% accuracy for single tree detection, and very strong correlations between measured AGB (above-ground biomass). The model was accurate, but also generalized well across different forest types. This made it suitable both for regional and national carbon accounting.
Model was tested at the Fangshan District in Beijing. This area has fragmented plantations, mainly Populus tomentosa and Pinus tabulaeformis. The AI model generated canopy height maps that closely matched ground truth data using Google Earth images with a resolution of one meter and references derived from lidar. The model outperformed CHM global products by capturing subtle differences in crown structure, which were often overlooked. Maps were generated that supported tree segmentation, plantation level biomass estimation and R2 values above 0.9. The network was also able to maintain its accuracy when it was applied to the forest of Saihanba. This confirmed that this system is adaptable across regions. Reconstructing annual growth patterns from satellite images archived over time is a solution that can be scaled for precision forestry and long-term monitoring of carbon sinks. This innovative technology bridges the divide between costly lidar surveys, and low-resolution optic methods. It allows detailed forest assessments with minimum data requirements.
“Our model demonstrates that large vision foundation models (LVFMs) can fundamentally transform forestry monitoring,” Dr. Xin Zhu, corresponding Author at Manchester Metropolitan University. “By combining global image pretraining with local self-supervised enhancement, we achieved lidar-level precision using ordinary RGB imagery. This approach drastically reduces costs and expands access to accurate forest data for carbon accounting and environmental management.”
The team used a deep learning framework that combined pre-trained LVFM with a process of self-supervised improvement. As inputs, high-resolution Google Earth images (2013-2010) were used. UAV lidar data was then used for validation and training. Model was trained on NVIDIA RTX GPUs using fastai and PyTorch. The model was compared to conventional networks such as U-Net, DPT and CHM global datasets. These experiments confirmed the superiority of this model in terms of accuracy and efficiency.
AI-based mapping offers an affordable and powerful approach to tracking forest growth and verifying credits. The adaptability of the system across different ecosystems allows it to be used for monitoring global programs on reforestation and afforestation. In the future, research will expand this method to include natural forests and mixed forest, incorporate automated species classification and provide real-time monitoring of carbon. These intelligent and scalable mapping technologies could be a key tool in the fight against climate change as we move towards net-zero.
Referrals
DOI
10.34133/remotesensing.0880
Original Source URL
https://spj.science.org/doi/10.34133/remotesensing.0880
Funding information
The National Science Foundation of China (72140005), and the Natural Science Foundation of Beijing (China) (grant number. This study is supported by the National Science Foundation of China (72140005) and Natural Science Foundation of Beijing, China (grant no. 2024NC-YBXM-220).
About Journal
Journal of Remote Sensing
Chuanlink Innovations is where innovative ideas are brought to life. The name Chuanlink Innovations, which is rooted in transmission and connection in its essence, represents our commitment to innovation, and to helping ideas to reach their full potential.
Click Here to Visit Related Site:
http://chuanlink-innovations.com
# # #