Abstract: (173 Views)
Background: The increasing concentration of greenhouse gases, especially carbon dioxide (CO₂), is the main driver of global warming and climate change. Urban forests, as part of green infrastructure, play a vital role in absorbing and sequestering this carbon. However, accurate estimation of above-ground carbon storage in urban trees through traditional field methods is costly, time-consuming, and limited in scale. In contrast, the integration of remote sensing data and machine learning models provides a novel, rapid, and cost-effective approach for large-scale monitoring. The main objective of this research was to estimate the above-ground carbon storage of trees in Sari city using vegetation indices extracted from Sentinel-2 satellite images and to compare the performance of four machine learning algorithms: Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbors (kNN), and Generalized Linear Model (GLM).
Methods: This study was conducted within the three urban districts of Sari city (with a total approximate area of 2970 hectares). Using selective sampling, 150 sample plots (50 plots in each district) were collected (In boulevards with the transect method and in squares with the circular sample plot method.). In each plot, the Diameter at Breast Height (DBH) and height of all trees were measured, and their locations were recorded with GPS. Stem biomass was calculated using an allometric equation based on DBH, height, form factor (0.5), and wood density, and carbon storage was estimated by multiplying it by a factor of 0.47. A set of cloud-free Sentinel-2 satellite images (2021-2022 timeframe) was used to extract a wide range of vegetation indices, including common indices (e.g., NDVI, EVI), red-edge band-based indices (e.g., S2REP, REIP, NDRE), and spectral transformations (e.g., TCB, TCW) on the Google Earth Engine platform. After screening variables based on multicollinearity (removing variables with Pearson correlation >0.8), the data were prepared for modeling. The performance of RF, SVM, kNN, and GLM models was evaluated using 10-fold cross-validation, and their accuracy was assessed with the coefficient of determination (R²), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) metrics. The relative importance of variables was also calculated for each model. All analyses were performed in the R software environment.
Results: The mean above-ground carbon storage in district one of Sari city was significantly higher than in the other two districts. The Random Forest (RF) model performed remarkably better than other models, showing the highest accuracy and lowest error. The observed vs. predicted values plot for RF showed a close fit to the 1:1 line, and its residuals were symmetrically scattered around zero, indicating no systematic bias. Although the SVM model showed acceptable accuracy in some iterations, it was unstable and exhibited high dispersion in error metrics. The kNN and GLM models performed weaker and showed a strong tendency to underestimate higher carbon storage values. The examination of variable importance in all models indicated the key role of red-edge band-based indices. The S2REP index had the highest relative importance in predicting carbon storage across all four models. Subsequently, the REIP, NDRE, and EVI indices ranked next in importance.
Conclusion: This research demonstrated that integrating advanced spectral indices from Sentinel-2 (especially red-edge indices) with the Random Forest machine learning algorithm provides an accurate, stable, and efficient method for estimating and spatially monitoring above-ground carbon storage of trees in the complex and heterogeneous urban environment of Sari. The superiority of RF stems from its ability to model complex nonlinear relationships, handle outliers, and deliver generalizable results. The pivotal role of the S2REP index emphasizes the importance of spectral information related to chlorophyll content and advanced physiological structure of trees (captured by red-edge bands) compared to conventional greenness indices for carbon storage estimation. It is suggested that urban managers and planners use this framework as an operational tool for periodic monitoring of carbon sequestration potential in green spaces, identifying strengths and weaknesses, prioritizing the protection of mature trees (such as old plane trees), and evaluating the effectiveness of green space development projects towards carbon reduction. For future studies, integrating multi-source data (such as airborne LiDAR for vertical canopy structure, Sentinel-1 radar data, and micro-scale environmental variables), developing ensemble models, and adapting the model for different tree species are recommended to increase the accuracy and reliability of estimates at the metropolitan level.
Methods: This study was conducted within the three urban districts of Sari city (with a total approximate area of 2970 hectares). Using selective sampling, 150 sample plots (50 plots in each district) were collected (In boulevards with the transect method and in squares with the circular sample plot method.). In each plot, the Diameter at Breast Height (DBH) and height of all trees were measured, and their locations were recorded with GPS. Stem biomass was calculated using an allometric equation based on DBH, height, form factor (0.5), and wood density, and carbon storage was estimated by multiplying it by a factor of 0.47. A set of cloud-free Sentinel-2 satellite images (2021-2022 timeframe) was used to extract a wide range of vegetation indices, including common indices (e.g., NDVI, EVI), red-edge band-based indices (e.g., S2REP, REIP, NDRE), and spectral transformations (e.g., TCB, TCW) on the Google Earth Engine platform. After screening variables based on multicollinearity (removing variables with Pearson correlation >0.8), the data were prepared for modeling. The performance of RF, SVM, kNN, and GLM models was evaluated using 10-fold cross-validation, and their accuracy was assessed with the coefficient of determination (R²), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE) metrics. The relative importance of variables was also calculated for each model. All analyses were performed in the R software environment.
Results: The mean above-ground carbon storage in district one of Sari city was significantly higher than in the other two districts. The Random Forest (RF) model performed remarkably better than other models, showing the highest accuracy and lowest error. The observed vs. predicted values plot for RF showed a close fit to the 1:1 line, and its residuals were symmetrically scattered around zero, indicating no systematic bias. Although the SVM model showed acceptable accuracy in some iterations, it was unstable and exhibited high dispersion in error metrics. The kNN and GLM models performed weaker and showed a strong tendency to underestimate higher carbon storage values. The examination of variable importance in all models indicated the key role of red-edge band-based indices. The S2REP index had the highest relative importance in predicting carbon storage across all four models. Subsequently, the REIP, NDRE, and EVI indices ranked next in importance.
Conclusion: This research demonstrated that integrating advanced spectral indices from Sentinel-2 (especially red-edge indices) with the Random Forest machine learning algorithm provides an accurate, stable, and efficient method for estimating and spatially monitoring above-ground carbon storage of trees in the complex and heterogeneous urban environment of Sari. The superiority of RF stems from its ability to model complex nonlinear relationships, handle outliers, and deliver generalizable results. The pivotal role of the S2REP index emphasizes the importance of spectral information related to chlorophyll content and advanced physiological structure of trees (captured by red-edge bands) compared to conventional greenness indices for carbon storage estimation. It is suggested that urban managers and planners use this framework as an operational tool for periodic monitoring of carbon sequestration potential in green spaces, identifying strengths and weaknesses, prioritizing the protection of mature trees (such as old plane trees), and evaluating the effectiveness of green space development projects towards carbon reduction. For future studies, integrating multi-source data (such as airborne LiDAR for vertical canopy structure, Sentinel-1 radar data, and micro-scale environmental variables), developing ensemble models, and adapting the model for different tree species are recommended to increase the accuracy and reliability of estimates at the metropolitan level.
Send email to the article author
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |