Regional Forest Carbon Stock Estimation Based on Multi-Source Data and Machine Learning Algorithms

Mingwei Zheng, Qingqing Wen, Fengya Xu, Dasheng WuForests 2025 · DOI

TL;DR

This study presents a refined method for estimating forest carbon stock (FCS) by integrating diverse remote sensing data (optical and radar) with crucial ecological and ground-based information. By employing advanced machine learning models and sophisticated feature selection, the research demonstrates that incorporating ecological features significantly boosts estimation accuracy, with LightGBM emerging as the top-performing algorithm.

What Problem Does This Paper Address?

Accurately quantifying forest carbon stocks is vital for understanding global carbon cycles, monitoring climate change mitigation efforts, and informing forest management strategies. Traditional methods, like field sampling, are labor-intensive and provide limited spatial coverage. While remote sensing offers a broader perspective, single data sources have limitations (e.g., cloud cover for optical, limited structural detail for radar). Furthermore, existing models often overlook critical ecological factors that influence carbon storage. This paper aims to bridge this gap by developing a more accurate and comprehensive FCS estimation model that leverages multi-source data and incorporates key ecological variables.

Methods

The core of this research lies in its multi-faceted approach to FCS estimation:

  • Multi-Source Data Integration: The study combines data from Sentinel-1 (SAR), Sentinel-2 (optical), Landsat 8 (optical), and Digital Elevation Models (DEM). This diverse data suite captures various aspects of forest structure and characteristics.
  • Ecological Feature Inclusion: Beyond spectral and structural data, the model incorporates four key ecological features: forest composition, primary tree species, humus thickness, and slope direction. These factors are known to influence carbon storage but are often underutilized in remote sensing models.
  • Feature Selection: To manage the large number of potential input variables and identify the most impactful ones, two feature selection techniques were employed:
  • Lasso: A regularization technique that can shrink coefficients of less important features to zero, effectively performing feature selection.
  • Recursive Feature Elimination (RFE): An iterative process that removes features based on their importance until a desired number is reached.
  • Machine Learning Models: Three powerful machine learning algorithms were used for FCS estimation:
  • XGBoost: An advanced gradient boosting algorithm known for its performance and regularization capabilities.
  • Random Forest (RF): A bagging ensemble method that builds multiple decision trees and aggregates their predictions.
  • LightGBM: A gradient boosting framework optimized for large-scale datasets, known for its speed and efficiency.

Datasets

The study utilized the following datasets:

  • Ground Data: Forest management and planning (FMP) data from the Lishui Forestry Bureau (2017), containing 38,994 samples of forest sub-compartments.
  • Digital Elevation Model (DEM): ASTER GDEM (version 2) with 30m resolution.
  • Optical Remote Sensing Imagery:
    • Sentinel-2 (S2) imagery (L2A product) acquired on December 25, 2017.
  • Landsat 8 (L8) imagery (L1TP product) acquired on November 3, 2017.
  • Radar Remote Sensing Imagery:
    • Sentinel-1 (S1) imagery (IW GRD-level data) acquired on December 10, 2017, using VV and VH polarization.

The study created a unified dataset of 36,187 valid samples, with 80% allocated for training and 20% for testing.

Key Insights and Ideas

  • Ecological Features are Crucial: The most significant finding is that incorporating ecological information (forest composition, primary tree species, humus thickness, slope direction) dramatically improves FCS estimation accuracy. This highlights the importance of moving beyond purely spectral and structural data.
  • LightGBM Dominates: Among the tested ML models, LightGBM consistently outperformed XGBoost and RF across various metrics, demonstrating its superior ability to handle complex forest data.
  • Feature Selection Matters: Both Lasso and RFE were effective in reducing the feature set, but they selected different subsets of variables. This suggests that the choice of feature selection method can influence model performance and interpretability depending on the data's characteristics.
  • Synergy of Data Sources: The combination of optical and radar data, along with DEM and ground truth, provides a more comprehensive understanding of forest structure and characteristics than any single source alone.

Key Results

The study developed four data schemes based on feature selection methods (RFE vs. Lasso) and the inclusion of ecological information. The results consistently showed that:

  • LightGBM is the top performer: In scheme B (RFE + ecological features), LightGBM achieved an R² of 0.78, MSE of 0.85, RMSE of 0.92, MAE of 0.58, rRMSE of 41.37%, and MAPE of 30.72%.
  • Ecological features boost performance: Schemes B and D (with ecological features) consistently outperformed schemes A and C (without ecological features) across all models and metrics. For instance, LightGBM's R² improved from 0.75 (Scheme A) to 0.78 (Scheme B), and MAPE decreased from 34.77% to 30.72%.
  • RFE vs. Lasso: While both feature selection methods improved model performance compared to using all features, RFE (especially when combined with ecological data in Scheme B) led to slightly better results than Lasso (Scheme D) in some aspects, though LightGBM's performance was strong in both.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Regional Applicability: The model was developed and validated for Qingyuan County, China. Its applicability to regions with different forest types, climates, and ecological conditions needs further validation.
  • Texture Feature Exploration: While texture features from SAR were used, exploring different window sizes, asynchronous lengths, and combining them with optical texture features could further enhance accuracy.
  • Temporal Resolution: The study primarily used single-temporal phase remote sensing images (autumn/winter). Acquiring data during the active tree growth period could provide more accurate vegetation information.
  • Multi-Temporal Data: Utilizing multi-temporal remote sensing data would capture seasonal and interannual dynamics, improving the model's spatial and temporal resolution and providing insights into long-term carbon stock trends and climate sensitivity.
  • Classification Accuracy: The accuracy of forest type and species classification, which influences the input ecological features, is critical and could be further improved.
  • Atmospheric Correction Consistency: Different atmospheric correction methods for various satellite data might introduce discrepancies. Exploring unified methods or comparative experiments is suggested.
  • Integration with Field Surveys: Combining remote sensing with field surveys and more accurate climate data could further refine assessment precision.

Summary updated

Abstract

Accurately assessing forest carbon stock (FCS) is essential for analyzing its spatial distribution and gauging the capacity of forests to sequester carbon. This research introduces a novel approach for estimating FCS by integrating multiple data sources, such as Sentinel-1 (S1) radar imagery, optical images from Sentinel-2 (S2) and Landsat 8 (L8), digital elevation modeling (DEM), and inventory data used in forest management and planning (FMP). Additionally, the estimation of FCS incorporates four key ecological features, including forest composition, primary tree species, humus thickness, and slope direction, to improve the accuracy of the estimation. Subsequently, insignificant features were eliminated using Lasso and recursive feature elimination (RFE) feature selection techniques. Three machine learning (ML) models were employed to estimate FCS: XGBoost, random forest (RF), and LightGBM. The results show that the inclusion of ecological information features improves the performance of the models. Among the models, LightGBM achieved superior performance (R2 = 0.78, mean squared error (MSE) = 0.85, root mean squared error (RMSE) = 0.92, mean absolute error (MAE) = 0.58, relative RMSE (rRMSE) = 41.37%, and mean absolute percentage error (MAPE) = 30.72%), outperforming RF (R2 = 0.76, MSE = 0.93, RMSE = 0.97, MAE = 0.60, rRMSE = 43.42%, and MAPE = 30.85%) and XGBoost (R2 = 0.77, MSE = 0.90, RMSE = 0.95, MAE = 0.61, rRMSE = 42.66%, and MAPE = 34.61%).
BibTeX
@inproceedings{zheng2025regional,
  author = {Mingwei Zheng and Qingqing Wen and Fengya Xu and Dasheng Wu},
  title = {{Regional Forest Carbon Stock Estimation Based on Multi-Source Data and Machine Learning Algorithms}},
  booktitle = {Forests},
  year = {2025},
  doi = {10.3390/f16030420},
}
Figure