A Novel Large Vision Foundation Model-Based Approach for Generating High-Resolution Canopy Height Maps in Plantations for Precision Forestry Management

Plantations for producing profitable products play a critical role in supporting local livelihoods, which requires accurate monitoring to guide sufficient management activities. Additionally, recent expansion of plantations for sequestering atmospheric CO 2 motivated by China Certified Emission Reduction (CCER) guideline further underscore the urgent need for accurate and cost-effective methods to estimate plantation aboveground biomass (AGB). High-resolution canopy height maps (CHMs) are essential for capturing detailed plantation traits for AGB estimation, especially given the typically small scale of plantations. While airborne or unmanned aerial vehicle (UAV)-based lidar remains the gold standard for acquiring high-resolution CHMs, its high cost limits widespread use. With advancements in deep learning, predicting CHMs using remote sensing RGB data has emerged as a cost-effective alternative, although challenges remain in accurately extracting canopy height-related features. To address these challenges, we develop a novel model for high-resolution CHM generation based on large vision foundation model (LVFM). This model integrates a feature extractor, a self-supervised feature enhancement module to avoid spatial detail loss in feature extraction, and a height estimator to produce high-resolution CHMs. Tested in the Fangshan District of Beijing, China—a region characterized by small, fragmented plantation parcels—our model, utilizing high-resolution (1-m grid) RGB imagery from Google Earth, demonstrated superior performance compared to existing methods, including conventional convolutional neural networks (CNNs) and naive LVFM implementations. The model achieved a mean absolute error of 0.09 m, a root mean square error of 0.24 m, and a correlation coefficient of 0.78 when evaluated against lidar-based CHM observations in pixel-wise assessments. Our model also exhibits satisfactory performance when being generalized into nontraining regions. Additionally, the CHMs generated by our model enabled over 90% success in individual tree detection and showed high accuracy in AGB estimation and a reasonable performance in tracking plantations’ growth. Our approach offers a promising tool for evaluating carbon sequestration in plantations and natural forests covering a large region.