The Aboveground Biomass Estimation of the Grain for Green Program Stands Using UAV-LiDAR and Sentinel-2 Data

Gaoke Yueliang, Gentana Ge, Xiaosong Li, Cuicui Ji, Tie‐Gang Wang, Tong Shen, Yubo Zhi, Chaochao Chen, Licheng ZhaoSensors 2025 · DOI

TL;DR

This study tackles the challenge of accurately estimating aboveground biomass (AGB) in China's "Grain for Green" (GGP) program's sparse, young forests. By integrating high-resolution UAV-LiDAR data with Sentinel-2 imagery and employing a Gradient Boosting Decision Tree (GBDT) model, the researchers developed a robust method for AGB estimation, achieving high accuracy and identifying key vegetation indices for monitoring these crucial ecological restoration areas.

What Problem Does This Paper Address?

The "Grain for Green" (GGP) program is a massive ecological restoration initiative in China aimed at combating land degradation. A critical aspect of evaluating its success is accurately measuring the aboveground biomass (AGB) of the reforested areas, as AGB is a direct indicator of forest health, carbon sequestration capacity, and overall ecosystem services. However, GGP stands, particularly in regions like Inner Mongolia, are characterized by sparse canopy cover, younger tree ages, and shorter tree heights. These features make traditional remote sensing methods, which often perform best in dense, mature forests, less effective. Existing research has been limited in addressing the specific challenges posed by these sparse, restoration-focused forest types, creating a need for specialized methodologies to accurately assess their AGB.

Methods

The study employed a multi-stage approach:

  1. AGB Sample Construction:
    • Individual Tree Segmentation: UAV-LiDAR point cloud data was processed to segment individual trees and extract their heights and locations.
  2. Optimal Growth Models: Field-measured tree height and Diameter at Breast Height (DBH) data were used to fit and select the best tree height-DBH growth models for major tree species (poplar, Mongolian Scots pine, Chinese red pine).
  3. DBH Estimation: The selected optimal models, combined with LiDAR-derived tree heights, were used to estimate DBH for individual trees.
  4. AGB Calculation: Species-specific biomass equations (based on national standards) were applied to estimate the AGB of individual trees.
  5. Sample Dataset Creation: LiDAR data was spatially gridded (30m resolution), and the AGB of all trees within each grid cell was summed to create a high-quality AGB sample dataset. Samples at stand edges or near roads/buildings were removed to ensure data quality.
  6. AGB Estimation Model:
    • Variable Extraction: A comprehensive set of 26 variables was derived from Sentinel-2 imagery (spectral bands, vegetation indices like NDVI, SAVI, and tasseled cap transformation indices like TCG) and terrain data (DEM, aspect, slope).
  7. Gradient Boosting Decision Tree (GBDT): A GBDT regression model was trained using 80% of the AGB sample dataset. The model aimed to predict AGB based on the extracted variables.
  8. Model Optimization: Key GBDT parameters (number of weak learners and learning rate) were optimized using a grid search method to maximize the R² on a test set.
  9. Prediction: The trained GBDT model was applied to Sentinel-2 imagery across the study area to generate an AGB distribution map.

Datasets

  • UAV-LiDAR Data: High-precision point cloud data collected from three locations in Inner Mongolia, covering approximately 370 hectares. This data was crucial for individual tree segmentation and height extraction.
  • Sentinel-2 Data: Zenith reflectance data from the Sentinel-2 satellite (Level-2A product) was used for vegetation index and tasseled cap transformation index calculations. Data from June 1 to September 1, 2023, was utilized.
  • Field Survey Data: Ground truth data collected from 79 sample plots across six county-level regions in Inner Mongolia during 2023-2024. This included measurements of tree height, DBH, crown width, and tree counts, which were essential for developing growth models and validating the AGB sample dataset.
  • SRTM DEM Data: 30m resolution Shuttle Radar Topography Mission Digital Elevation Model data was used to extract topographic variables.
  • GGP Stands Distribution Data: Vector data provided by the General Station of Forestry and Grassland Work of Inner Mongolia, detailing the location, boundaries, tree species, and implementation year of GGP stands.

Key Insights and Ideas

  • LiDAR as a Bridge: UAV-LiDAR data proved invaluable for bridging the gap between ground measurements and satellite imagery. Its ability to extract detailed 3D structural information (like tree height) was critical for building accurate AGB samples in sparse forests.
  • Optimized Growth Models Matter: Developing species-specific optimal tree height-DBH growth models was a crucial step. This highlights that generic models are insufficient for accurate biomass estimation, especially in diverse forest types.
  • Tasseled Cap Transformation is Key for Sparse Vegetation: The study strongly emphasizes the importance of the Tasseled Cap Greenness (TCG) index, along with other vegetation indices (RENDVI, VDVI), for AGB estimation in arid and semi-arid regions with sparse vegetation. TCG effectively captures subtle greenness variations that might be missed by traditional indices.
  • GBDT's Robustness: The GBDT algorithm demonstrated high accuracy and robustness in estimating AGB, effectively integrating diverse data sources (spectral, vegetation indices, terrain) to capture complex relationships.
  • AGB Sample Dataset as Ground Truth: The AGB sample dataset constructed using LiDAR data showed excellent consistency with field data (R² = 0.89), validating its use as a reliable proxy for ground truth in model development.

Key Results

  • AGB Sample Dataset Validation: The constructed AGB sample dataset achieved an R² of 0.89 and an RMSE of 781 g/m² when compared to field measurements, demonstrating its reliability.
  • AGB Estimation Model Accuracy: The GBDT model achieved a high validation accuracy with an R² of 0.96 and an RMSE of 560 g/m².
  • Important Variables: Tasseled Cap Greenness (TCG), Red-Edge Normalized Difference Vegetation Index (RENDVI), and Visible-band Difference Vegetation Index (VDVI) were identified as the most influential variables for AGB estimation.
  • AGB Ranges: The estimated AGB for major tree species in the new round of GGP stands in Inner Mongolia ranged from 120 to 9253 g/m².
  • Poplar: Mean AGB of 978 g/m².
  • Mongolian Scots pine: Mean AGB of 622 g/m².
  • Chinese red pine: Mean AGB of 313 g/m².
  • Regional Differences: The eastern region of Inner Mongolia had a higher average AGB (757 g/m²) compared to the western region (465 g/m²).

Limitations and Future Work

  • Sentinel-2 Resolution: The study notes that the spatial resolution of Sentinel-2 is limited for capturing subtle AGB variability in very sparsely vegetated areas of GGP.
  • Future Work Suggestions:
    • Higher Resolution Data: Incorporating higher-resolution remote sensing data (e.g., from GF-1 imagery) or combining multispectral data with UAV-LiDAR in future studies could further enhance AGB estimation capabilities.
  • Dynamic Monitoring: The developed methodology provides a solid foundation for long-term dynamic monitoring of GGP stands.
  • Broader Application: Extending this approach to other regions with similar sparse forest characteristics could be beneficial.

Summary updated

Abstract

Aboveground biomass (AGB) serves as a crucial indicator of the effectiveness of the Grain for Green Program (GGP), and its accurate estimation is essential for evaluating forest health and carbon sink capacity. However, due to the dominance of sparse forests in GGP stands, research in this area remains significantly limited. In this study, we developed the optimal tree height-diameter at breast height (DBH) growth models for major tree species and constructed a high-quality AGB sample dataset by integrating airborne LiDAR data and tree species information. Then, the AGB of the GGP stands was estimated using the Sentinel-2 data and the gradient boosting decision tree (GBDT) algorithm. The results showed that the AGB sample dataset constructed using the proposed approach exhibited strong consistency with field-measured data (R<sup>2</sup> = 0.89). The GBDT-based AGB estimation model shows high accuracy, with an R<sup>2</sup> of 0.96 and a root mean square error (RMSE) of 560 g/m<sup>2</sup>. Key variables such as tasseled cap greenness (TCG), red-edge normalized difference vegetation index (RENDVI), and visible-band difference vegetation index (VDVI) were identified as highly important. This highlights that vegetation indices and tasseled cap transformation index information are key factors in estimating AGB. The AGB of major tree species in the new round of the GGP stands in Inner Mongolia ranged from 120 to 9253 g/m<sup>2</sup>, with mean values of 978 g/m<sup>2</sup> for poplar, 622 g/m<sup>2</sup> for Mongolian Scots pine, and 313 g/m<sup>2</sup> for Chinese red pine species. This study offers a practical method for AGB estimation in GGP stands, contributing significantly to sustainable forest management and ecological conservation efforts.
BibTeX
@inproceedings{yueliang2025aboveground,
  author = {Gaoke Yueliang and Gentana Ge and Xiaosong Li and Cuicui Ji and Tie‐Gang Wang and Tong Shen and Yubo Zhi and Chaochao Chen and Licheng Zhao},
  title = {{The Aboveground Biomass Estimation of the Grain for Green Program Stands Using UAV-LiDAR and Sentinel-2 Data}},
  booktitle = {Sensors},
  year = {2025},
  doi = {10.3390/s25092707},
}
Figure