نوع مقاله : علمی - پژوهشی
عنوان مقاله English
نویسندگان English
Urban development and the need for carbon resource management have increased the importance of accurately estimating urban tree biomass. LiDAR technology is considered an efficient tool for modeling tree canopy structure. However, the complexity of urban scenes, including interference from buildings and other artificial features, has made the automatic extraction of individual trees challenging. The main objective of this study is to present an integrated machine learning-based framework for accurate point cloud segmentation, individual tree crown extraction, and biomass estimation in urban environments. The main innovation of this study is the presentation of a hybrid framework combining deep learning and machine learning which, unlike previous studies, without requiring high-density training data and independent of prior knowledge of tree species, is capable of estimating biomass in complex urban environments using only medium-density LiDAR data. The data used included LiDAR point clouds of the New York urban area with an average density of 4 to 10 points per square meter and reference field measurements including trees of 9 different species. The DALES dataset with a density of 50 points per square meter was also used as auxiliary training data. In the first step, the PointCNN neural network was used to classify point clouds into three classes: ground, building, and other features, and the Early Stopping technique was used to prevent overfitting. Another innovation is that, unlike common approaches that apply the deep network to all classes, in this study, due to the density difference between the training data and the main data, PointCNN was used only to detect ground and building points, and tree point extraction was performed through a complementary processing framework. This framework includes removal of building points, generation of a digital terrain model and canopy height model, height normalization, removal of noise points, formation of convex hulls using the Alpha Hull method, removal of unsuitable polygons, application of a circularity criterion, generation of a canopy surface model, smoothing of the canopy height model, identification of local maximum points, and segmentation with the Watershed Segmentation algorithm. For each crown, geometric features including tree height at the 10th to 90th percentiles, longest diameter of the inscribed ellipse, ellipse area, and geometric volume of trees were calculated. Three machine learning algorithms—Random Forest, LightGBM, and Support Vector Regression—were evaluated for modeling the relationship between features and biomass. Hyperparameter tuning was performed with Bayesian optimization and five-fold cross-validation. Three machine learning algorithms, including Random Forest, LightGBM, and Support Vector Regression, were evaluated for modeling the relationship between features and biomass. The PointCNN model, with an F1-score of 0.98 for the ground class and 0.91 for the building class, showed high performance in feature separation. Segmentation after smoothing reduced local maximum points to one point per tree crown. The minimum acceptable tree crown area was set at 3 m². Evaluation of the regression models showed that Random Forest had the best performance, with an RMSE of 464.14 kg and R² of 0.77 and a slope close to one in the scatter plot. The LightGBM algorithm had higher speed but greater error, and Support Vector Regression provided the weakest estimate. Feature importance analysis showed that the upper percentiles of tree canopy height, including H80, H60, and H90, are the most important variables. The final biomass map in 100×100 m cells showed high agreement with the green space condition of the region, such that the highest biomass was observed in the southwest (Dyker Beach Park) and the lowest in areas adjacent to the river. Despite challenges such as differences in data density and imbalance in tree species samples, the proposed framework was able to provide acceptable estimates of urban tree biomass. The main advantage of the method is its independence from prior knowledge of tree species and no need for aerial or satellite imagery. This study shows that LiDAR data with an average density of 7 points per square meter can be an efficient tool for monitoring urban carbon resources without the need for supplementary data.
کلیدواژهها English