Forest ecosystems are the largest terrestrial carbon sink, playing a crucial role in the global carbon cycle and climate change mitigation. Accurate estimation of forest carbon stocks is essential for building reliable carbon cycle models and supporting carbon neutrality policies. Traditional methods for dynamic carbon stock assessment are generally conducted at the stand scale and lack the resolution to capture variations at the individual tree level. This study systematically estimated the aboveground carbon stock growth of individual trees in poplar plantations. We investigated 12 poplar (Populus L.) plantation plots with varying planting spacings and clones, using high-resolution terrestrial laser scanning (TLS) data acquired in 2019 and 2021. Several basic tree measuring factors were extracted from branch skeletons based on a novel algorithm of the incomplete simulation of tree transmitting water and nutrients (ISTTWN). After correlation analysis and multicollinearity diagnostics, the Boruta algorithm was applied to select significant predictors for modeling. Both linear and nonlinear approaches were compared, along with four machine learning models: random forest, k-nearest neighbors, support vector machine, and CatBoost. Hyperparameters were optimized using the Optuna framework with mean squared error (MSE) as the objective function in 5-fold cross-validation. The best-performing model was selected to estimate individual tree aboveground carbon stock and carbon stock growth. Two estimation approaches were evaluated: a direct method that modeled carbon stock growth as the dependent variable, and an indirect method that derived growth from the difference between carbon stocks estimated at two time points. The more accurate approach was identified for estimating aboveground carbon stock growth at the individual tree level, and the optimal planting configuration was determined for poplars in the study area. Correlation analysis indicated that both the individual tree aboveground carbon stock and carbon stock growth showed strong correlations with diameter at breast height (DBH) and tree volume (V). Among crown structural variables, crown surface area (CSA) exhibited the strongest correlation. Due to the significant multicollinearity among the basic tree measuring factors, variables with high variance inflation factor (VIF) values were excluded. Consequently, tree height (H), branch number (N), average branch length (BLa), average branch diameter (BDa), and average depth into crown (DINCa) were ultimately selected for constructing linear and nonlinear models to estimate carbon stock. For modeling carbon stock growth, the selected predictors were tree height (H), crown width (CW), branch number (N), average branch diameter (BDa), and average depth into crown (DINCa). Results showed that the random forest model with Boruta-selected variables outperformed other models, achieving an R2 of 0.944 for aboveground carbon stock and 0.798 for carbon stock growth. Specifically, the direct estimation approach using the random forest model yielded the most accurate predictions for growth, with an R2 of 0.821, a root mean square error (RMSE) of 0.920 kg, and a mean absolute error (MAE) of 0.733 kg. These results demonstrate the high precision and robustness of machine learning in capturing complex nonlinear relationships between basic tree measuring factors and carbon dynamics. Furthermore, the NL-797 poplar clone planted at a spacing of 6 m × 6 m exhibited the highest absolute carbon stock growth per tree, along with a consistently high growth rate, indicating that this configuration is favorable for carbon accumulation in poplars. This highlights how specific genotypes combined with optimal planting density can significantly enhance carbon sequestration in plantation forests. In conclusion, this study presents a non destructive and accurate method for estimating aboveground carbon stock growth of individual trees by integrating TLS data and advanced machine learning modeling. The findings offer valuable insights for forest management strategies aimed at maximizing carbon storage and provide a scientific basis for designing carbon oriented plantation forests.
- Article type
- Year
- Co-author
In order to explore the selection and number of modeling variables in biomass models and their impact on the final fitting accuracy, providing methodological references for the establishment of biomass models.
Taking Eucalyptus plantations as the research object, CHM was constructed using unmanned aerial vehicle (UAV) airborne LiDAR point cloud data. Gaussian low-pass filtering and enhanced Frost filtering were applied to calculate the arithmetic mean height of the sample plots. Using variable projection importance method and VSURF package to screen variables to compare the differences in biomass fitting effects between multiple regression models and machine learning models, and select the optimal model for subsequent research.
1) The arithmetic mean of the sample plots after enhanced Frost filtering and Gaussian low-pass filtering is higher, and its relative error is lower than the arithmetic mean of the sample plots directly extracted on CHM. Among them, the extraction effect of enhanced Frost filtering is slightly better than that of Gaussian low-pass filtering; 2) The fitting accuracy of multiple regression models increases with the increase of variables, and nonlinear models generally outperform linear models. In machine learning, the random forest model performs the best, with R2 of 0.88, RMSE of 16.15 t/hm2, and MAE of 12.17 t/hm2, and the fitting effect is better than that of multiple regression models; 3) The variables filtered using the VSURF package have a better modeling effect compared to directly using all variables. After variable screening, it was found that the height feature variables of point clouds have a higher importance on biomass compared to density and intensity variables, indicating that height variables have a stronger explanatory power on forest biomass.
Using airborne LiDAR point cloud data, enhancing Frost filtering to smooth images can significantly reduce the error in extracting tree height, and the extraction effect is slightly better than Gaussian low-pass filtering. Using the VSURF package to filter variables can improve the accuracy of the model, and the random forest model performs best in estimating the biomass of eucalyptus trees in artificial forests.
This work aimed to explore the spatial distribution of forest carbon storage based on geographic code, to provide a method reference for the long-term spatial change analysis of forest carbon storage.
Based on the Second Survey data of Forest Resources of the Purple Mountain in 1987, 2002 and 2019, by taking advantage of the biomass conversion factor continuous function method and the geocoded fixed grid, the study analyzed the dynamics of forest carbon storage and its spatial distribution changes of the Purple Mountain in the past 32 years.
1) In terms of time dimension, the forest carbon storage in the study area in 1987, 2002 and 2019 were 62 541.1 t, 81 703.0 t and 106 281.8 t, respectively, with a net increase of 43 740.7 t and an average annual increase of 1 325.5 t in the past 32 years. The carbon storage of broad-leaved arbor forests accounted for 62.1%, 73.0% and 81.7% of the forest carbon storage, which was significantly greater than that of coniferous arbor forests and coniferous broad-leaved mixed arbor forests. Between 1987 and 2002, the sum of the carbon storage of middle-aged forests and mature forests accounted for 76.4% and 60.4% of that of arbor forests. In 2019, the carbon storage of mature and over-mature forests accounted for 59.5% of the total carbon storage of arbor forests. 2) In terms of spatial distribution, the forest carbon storage in the study area showed a gradually increasing spatial distribution from south to north. The regional proportion of medium carbon storage and high carbon storage increased from 45.1% in 1987 to 53.9% in 2002 and 60.3% in 2019. The stable, medium and high stability areas accounted for 71.1% of the study area, and generally, the stable area of forest carbon storage accounted for a relatively high proportion. 3) In terms of the role of forest management in the space-time dimension, in view of the difficulties in analyzing the long-term spatial change of regional forest carbon storage and the low accuracy of the clear spatial location of forest management, the method proposed in this paper can not only compare the distribution characteristics of forest carbon storage in the long-time series of the same spatial location but also compare and analyze the forest management effects of carbon sequestration and carbon increase in the specific spatial location, which is also convenient for query and supervision. Meanwhile, it can provide support for the long-term grid management of forest head systems and offer technical support for the accurate improvement of regional forest carbon sequestration management.
The spatial distribution of forest carbon storage has been studied by using the fixed grid of geocoding and the coefficient of variation index, which provides a reference for the sustainable monitoring of forest carbon storage in the Purple Mountain in the future.
As an important unit of forest stock estimation, standing wood volume has important significance in forest resource investigation. This paper studied the method of obtaining the attribute factors of standing wood branches based on multi-source Lidar, and explored the ability of tree point cloud to build a better predicting model of standing wood volume.
In this paper, based on the fusion point cloud data of ground-based Lidar (TLS) and airborne Lidar (ALS), a three-dimensional single tree model was established by using the geometric characteristics of tree skeleton and the extraction algorithm of incomplete simulation of water and nutrient transport (ISTTWN), and the branch attribute factors of individual poplar trees were obtained. By constructing a prediction model of stand volume with branch attribute factor as independent variable, the optimal estimation model of stand stock was explored.
The accuracy of branch attribute factors after fusion was much improved compared with that before fusion, and the extraction accuracy was in the order of branch height > branch length > chord length > branch growth angle > branch angle > bow height. Among them, the fit degree of branch length was the highest with R2 of 0.989. Compared with the linear and nonlinear product volume models established by the feature parameters, the model constructed based on the feature parameters increased by 0.088 and 0.110 respectively, while the RMSE decreased by 0.012 and 0.009 m3 respectively. The linear and nonlinear models fit 0.688 and 0.709 respectively, which was the best among the six groups of volume prediction models.
After the fusion of point cloud data between TLS and ALS, the high point cloud density can be effectively improved due to the mutual compensation between the data, and the extraction accuracy of branch attribute factors can be significantly improved in the research and development of 3D tree models. At the same time, adding the independent variable of branch attribute factor into the volume prediction model can effectively improve the accuracy of the model prediction.
京公网安备11010802044758号