| Abstrakt: |
Background and Objectives: Estimating forest tree structural attributes such as height and diameter at breast height (DBH) is crucial for understanding the structure and management of forest resources. One important method for estimating these parameters is the individual tree detection (ITD) method using appropriate remote sensing data, such as airborne LiDAR data. However, it should be noted that different ITD methods have various limitations and capabilities and react differently to changes in forest tree species and the vertical structure of the canopy. Methodology: This study presents a hybrid individual tree detection method that combines raster-based and point-based methods in a multi-scale framework to identify single trees from LiDAR data. In this method, tree crown scale levels are obtained from morphological filters in the canopy height model (CHM). Segmentation is then performed using a multi-scale method, and the results are merged. To better separate adjacent and understory trees, the point cloud inside the segments is analyzed using the probability density function, and tree crown segments are modified. After detecting single trees, DBH and height parameters were estimated using ground control data and extracted features from LiDAR data with machine learning algorithms, including random forest (RF), support vector machine (SVM), and cubist (CB), in the form of 10-fold nested cross- validation (10-fold NCV). The Boruta feature selection algorithm was used to identify the most important metrics based on the LiDAR point cloud, which played an effective role in improving the performance of machine learning algorithms. Due to limited access to LiDAR and ground data from Iran’s forests, this study uses the NEWFOR single tree detection benchmark dataset, collected from forests of the Alpine region with a combination of different tree species and vertical canopy structures. Results: Although understory trees cannot be extracted with the same accuracy as overstory trees, the results of this study showed that, on average, the developed multi-scale individual tree detection (MSITD) method detected 89% of the tree crowns in the highest height layer and the highest number of small overstory trees with a detection rate of 48% in the lowest height layer (2-5 meters). The analysis of the machine learning algorithms’ results in estimating forest structural attributes showed that, despite slight differences in performance, the SVM algorithm performed better than the RF and CB algorithms in estimating both height and DBH attributes. For the height attribute, the mean values of RMSE, rRMSE, and R² in the SVM algorithm were 1.75 m, 9%, and 0.85, respectively. For the DBH attribute, the values obtained for RMSE, rRMSE, and R² were 4.74 cm, 19%, and 0.78, respectively. Conclusion: The evaluation of the results showed that the methods presented in this study for identifying single trees and estimating forest tree structural attributes have high potential for practical applications. [ABSTRACT FROM AUTHOR] |
