A Lightweight LiDAR SLAM System Integrated with Semantic Segmentation for Forest Biomass Parameter Acquisition

A Lightweight LiDAR SLAM System Integrated with Semantic Segmentation for Forest Biomass Parameter Acquisition | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article A Lightweight LiDAR SLAM System Integrated with Semantic Segmentation for Forest Biomass Parameter Acquisition Yunfeng Zhao, Shipeng Zhao, Yongxu Zhou, Bing Shen, Jing Wang This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6971887/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 11 You are reading this latest preprint version Abstract Context : Efficient and automated acquisition of individual tree parameters is essential for intelligent forest resource inventory. Traditional methods, which rely heavily on manual measurements, are limited in their scalability and cannot meet the demands of large-scale, high-frequency surveys. Aims : This study aims to develop a lightweight LiDAR SLAM system integrated with semantic segmentation to improve the accuracy and real-time performance of individual tree extraction and DBH measurement in forest environments. Methods: The system is based on an improved LIO-SAM framework, incorporating the IKD-Tree to enhance efficiency. In the semantic segmentation module, the SqueezeSegV3 network with the added ELA attention mechanism is employed to improve semantic category recognition. The segmented point clouds are processed through clustering and cylindrical fitting to extract individual trees and estimate their DBH. Results : On both a public dataset and field-collected forest data, the semantic segmentation achieved accuracies of 0.85 and 0.89, with mean Intersection over Union values of 0.55 and 0.67, respectively. The average prediction accuracy for DBH reached 98.6%.. Conclusion The system integrates a lightweight semantic network with an efficient point cloud structure, offering both high accuracy and real-time performance. It meets the requirements of large-scale and high-efficiency measurements in forest resource inventory tasks. 3D LiDAR Forest Point cloud Semantic segmentation Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Key message A lightweight SLAM system integrated with semantic segmentation is proposed, enabling efficient and accurate single-tree extraction and diameter at breast height (DBH) measurement in complex forest environments. 1. Introduction Aboveground forest biomass(tree height, DBH, and species composition)is a fundamental component and key indicator in forest inventory and assessment (Zhang et al. 2024,Mu et al. 2025). Conventional forest resource surveys are typically conducted through manual measurements, which demand substantial labor and time and may also negatively impact the natural environment(Gollob et al. 2021). Advances in sensor technologies have facilitated the wide-spread use of remote and proximal sensing techniques in automated forest surveys across multiple spatial scales(Tian et al. 2023,Bárta et al. 2022). However, due to its inability to capture the three-dimensional structure of forests, optical imagery is limited in providing com-prehensive biomass estimations. The accurate acquisition of forest structural details and key parameters is currently achieved primarily through Terrestrial Laser Scanning (TLS), Unmanned Aerial Vehicle Laser Scanning (UAV-LS), and Mobile Laser Scanning (MLS) (Xu et al. 2021b). Terrestrial laser scanning is time-consuming in data acquisition (Brede et al. 2019) and is ineffective in estimating individual tree parameters under occluded conditions(Bauwens et al. 2016), making it un-suitable for complex forest environments. UAV-LS, which is not constrained by forest terrain, offers advantages such as rapid data acquisition and ease of deployment, making it widely used for forest data collection. Neuville et al. (2021) utilized UAV-acquired LiDAR data and applied machine learning techniques to estimate forest structure. Similarly, Fadi et al. (2024) used UAV data combined with allometric equations to estimate forest biomass and carbon storage. Previous studies have demonstrated the effective-ness of UAV in forest data collection. However, since UAVs primarily operate above the canopy, foliage occlusion often prevents accurate measurements of sub-canopy forest structure and ground-level forest biometric parameters (Hao et al. 2020). In comparison, MLS sys-tems excel in ground-based coverage, enabling detailed 3D reconstruction of forest understory architecture and significantly improving single-tree feature extraction ac-curacy (Gao and Kan 2022). Mobile Laser Scanning is mainly divided into airborne and backpack-mounted systems. By integrating Simultaneous Localization and Mapping (SLAM) technology, MLS can effectively acquire large-scale forest data (Kukko et al. 2017). SLAM relying on its excellent real-time performance and high-precision mapping capabilities, has been widely ap-plied in forestry resource surveys (Fan et al. 2020,Gollob et al. 2020), Tang et al. (2015)investigated a LiDAR-IMU SLAM system mounted on an all-terrain vehicle (ATV) for forest inventory applications. Building upon this work, Qian et al. (2016) incorporated GNSS measurements into the system to achieve higher positioning accuracy. Pierzchała et al. (2018) utilized 3D LiDAR combined with SLAM to generate forest maps and estimate individual tree positions and diameters at breast height (DBH). Besides airborne LiDAR acquisition, back-pack-mounted laser scanning has also been widely employed. For example, Oveland et al. (2018) developed a backpack system equipped with two 3D LiDAR sensors featuring vertical and horizontal rotation axes. This dual-LiDAR configuration enables the system to cover a wider area during scanning and capture more feature information from multiple angles. Zhou et al. (2024) proposed a backpack-mounted laser scanning system with the F2-SLAM algorithm, which extracts ground and tree trunk features from LiDAR data and employs least squares adjustment to optimize LiDAR scan registration. The system also integrates trajectory from the navigation unit to enhance feature extraction. The above studies demonstrate that integrating SLAM technology with LiDAR for forest data collection enables fast, accurate, and comprehensive acquisition of the three-dimensional structure of trees and vegetation (Yang et al. 2024), thereby improving data col-lection efficiency and reducing the time and cost associated with manual measure-ments. However, due to the complex spatial structure of forests, relying solely on 3D LiDAR-based SLAM technology provides limited geometric information (Wang et al. 2022), making it difficult to handle scenarios with highly similar tree structures and uneven density distributions. For instance, densely clustered multi-stem trees may be mistakenly identified as a single treetop (Yu et al. 2024), resulting in poor segmentation accuracy (Rajab Pourrahmati et al. 2023). Addi-tionally, large treetop regions may be incorrectly detected as individual trees, leading to segmentation errors. These issues significantly affect the reliability of single-tree ex-traction. To address this challenge and improve segmentation accuracy, this study in-corporates deep learning to obtain semantic information of the forest, thereby providing the system with more robust feature representations. Deep learning has become an ef-fective solution for addressing the challenges of single-tree segmentation in complex forest environments. For example, Chen et al. (2020) proposed SLOAM, a semantic LiDAR SLAM system for forest resource inventory that can extract individual tree information through instance segmentation and automatically estimate DBH. However, this system is prone to interference from underbrush noise during ground point cloud segmentation. Liu et al. (2022) presented a LiDAR point cloud semantic segmentation method tailored for complex forest environments, capable of accurately identifying underbrush and individual trees, but the algorithm lacks spatial hierarchical awareness of forest structure, making it difficult to adapt to complex scenarios. Ma et al. (2023) introduced the Forest-PointNet model specifically designed to extract vertical forest structural information from terrestrial LiDAR data, achieving high-precision semantic segmentation of different structural layers in complex forest scenes. Li et al. (2023) pro-posed the Point DMM method, which combines structured annotation with multi-scale feature extraction to effectively improve segmentation accuracy and robustness. How-ever, its complex model architecture entails a high demand for computational re-sources. The aforementioned studies demonstrate the effectiveness and feasibility of deep learning techniques in acquiring tree parameters in forestry applications. Never-theless, achieving high-precision segmentation typically relies on large-scale point cloud datasets, which poses challenges to system efficiency and real-time performance, especially when deploying on embedded platforms such as unmanned vehicles. Therefore, developing a lightweight and real-time semantic SLAM system has become a current research focus. Such a system not only facilitates fine-grained classification of forest resources but also provides efficient and reliable data support for individual tree parameter extraction, stand structure analysis, and precise biomass estimation. To address the aforementioned issues, we propose a LiDAR SLAM-based forestry tree resource inventory system integrated with deep learning. The system primarily utilizes unmanned vehicles and LiDAR sensors to generate semantic point cloud maps of forests, enabling the measurement of DBH and tree count. To achieve more accurate and real-time output of forest internal structure information, we employ the V3 algo-rithm to process LiDAR point cloud data, extracting features such as trunks and ground through semantic segmentation. Compared with traditional point cloud clustering methods, such as LiDAR odometry and mapping (He and Li 2020),(Eeckhaut et al. 2007), the proposed system demonstrates superior accuracy and efficiency. The main contributions of this study include: propose a lightweight semantic LiDAR SLAM system that integrates an improved LIO-SAM algorithm with the semantic segmentation network SqueezeSegV3 to construct real-time 3D maps. By fusing semantic information with point clouds, the system enriches the point cloud with semantic features and enhances mapping accuracy. incorporate the incremental k-d tree (IKD-Tree) algorithm for dynamic point cloud updates, which significantly improves the processing speed of the SLAM system and ensures real-time performance in large-scale environments. introduce a semantic segmentation-based individual tree extraction method, which reduces the influence of noise on point cloud clustering and improves the accuracy of individual tree identification. 2. Material and methods 2.1. Hardware equipment The experimental platform used in this study for forest inventory is shown in Fig. 1 . The unmanned vehicle chassis measures 0.726 m in length, 0.617 m in width, and 0.273 m in height, with a maximum payload capacity of 50 kg. It adopts differential steering and has a maximum operating speed of 0.8 m/s, with a climbing ability of up to 25°. The vehicle supports manual control via a joystick, with a maximum control range of 20 meters.The vehicle is equipped with a MID360 LiDAR sensor, which collects 20,000 point cloud data points per second at a frequency of 10 Hz. The sensor features a 360° horizontal and a vertical viewing range from − 7° to 52°. It is managed by an integrated PC acting as the upper computer, powered by an AMD Ryzen3 3200G CPU (3.6 GHz). The system operates on Ubuntu 18.04 with the Robot Operating System (ROS) frame-work. 2.2. Study area The study area selected for this research is Lihu Forest Park, located in Gu'an County, Langfang City, Hebei Province. The park consists of a mix of natural forests and artificially planted woodlands. Two forest plots with different tree species within this area were selected as data collection sites. Each plot covers an area of 25m*25m. The collected data from each plot were partitioned and used for training and testing purposes. 2.3. System Overview In This study adopts LIO-SAM (Shan et al. 2020) as the foundational algorithm and makes im-provements tailored to the application scenario, as illustrated in the system framework shown in Fig. 3 . The system comprises three main components: a SLAM module, a semantic segmentation module, and an individual tree extraction module. After the LiDAR mounted on the unmanned vehicle acquires point cloud data, the data are first input into both the SLAM and semantic segmentation modules. In the SLAM module, to enhance frontend processing efficiency, the original downsampling algorithm is re-placed with the ikd-tree method (Cai et al. 2021), accelerating feature extraction and reducing computational load. The semantic segmentation module utilizes an improved Squeez-eSegV3 algorithm (Xu et al. 2021a) combined with a local attention mechanism, ELA (Xu and Wan 2024), effec-tively enhancing the recognition of features such as ground and tree trunks. The seg-mented point clouds are then fed back to the SLAM module to improve mapping ac-curacy. Meanwhile, the individual tree extraction module uses the segmented point clouds as the basis for subsequent processing. During individual tree extraction, the system obtains classified trunk and ground point clouds, applies the DBSCAN clus-tering algorithm to identify individual trees, and performs cylindrical fitting through linear fitting methods to estimate tree DBH. 2.4. Ikd-tree algorithm The K-D Tree is an efficient data structure that organizes multidimensional point data, enabling fast nearest neighbor searches (Nuchter et al. 2007). However, the K-D Tree is inherently static, requiring reconstruction and merging from scratch each time a new data frame is received. This process is time-consuming and negatively impacts system efficiency, making it unsuitable for lightweight and real-time applications. Therefore, we adopt the more efficient IKD-Tree algorithm to replace the original K-D Tree in our system. The IKd-Tree is a dynamic data structure based on the K-D tree, with its core composed of a binary search tree capable of incremental updates. Each node corresponds to the minimum bounding rectangle defined by its associated point set and stores the splitting axis along with the corresponding threshold to divide the node into two subregions. To support real-time performance, the IKd-Tree introduces incremental operations such as point insertion, reinsertion, and deletion. These operations are performed through recursive updates of local node states, and when necessary, local subtree reconstruction is triggered to maintain overall structural balance and query efficiency. In addition, the algorithm incorporates a lightweight dynamic rebalancing mechanism to prevent efficiency degradation caused by structural imbalance during construction and long-term operation. Specifically, when a subtree imbalance is detected, parts of the structure are selectively decomposed and locally reconstructed in parallel. This approach preserves accuracy while minimizing the impact of rebalancing on the system’s overall runtime performance. By leveraging localized reconstruction and multithreaded parallelism, the IKd-Tree maintains high responsiveness and efficient query performance even under high-frequency updates and large-scale point cloud processing scenarios. In the IKd-Tree, point insertion is performed by sequentially comparing coordi-nate values along the splitting axes to determine the appropriate position within the tree. Based on the local structural state of the tree, subtrees may be selectively recon-structed to maintain overall balance. Theoretically, in the worst case, an insertion op-eration may require up to 𝑂(𝑛)comparisons. However, thanks to dynamic maintenance strategies and point count control mechanisms, the actual insertion efficiency is typi-cally stabilized at logarithmic complexity in practice. For point deletion, the IKd-Tree similarly employs a recursive search to locate the target node, followed by local sub-tree reconstruction when necessary to maintain structural balance. During query oper-ations, the algorithm evaluates each level based on the splitting dimension and recur-sively traverses the left and right subtrees. In most cases, the query complexity re-mains at O(log n). By incorporating rebalancing mechanisms during both insertion and deletion stages, the IKd-Tree significantly enhances computational efficiency and structural stability in large-scale point cloud processing tasks. 2.5. semantic segmentation framework Semantic segmentation helps to compensate for the lack of feature information in LiDAR data under complex environments. In this study, we adopt the SqueezeSegV3 algorithm as the core semantic segmentation framework. This algorithm is a light-weight segmentation network based on range images and leverages spherical projec-tion to map 3D point clouds onto 2D spherical grids, resulting in dense range imag-es.To better handle the spatially varying feature distributions of LiDAR images, the algorithm introduces a Spatially-Adaptive Convolution (SAC) module. This module provides efficient spatial adaptability and content-awareness, making it well-suited for forest environments with complex spatial structures. The semantic segmentation workflow is illustrated in the Fig. 5 . To improve the performance of the segmentation algorithm in forest environ-ments, we integrate the Efficient Local Attention (ELA) mechanism into our frame-work. After obtaining the 2D range image, the image features are first passed through the ELA module, which accurately captures the location of regions of interest. This method maintains the original channel dimensions of the input feature maps and pre-serves the lightweight nature of the network.The local attention mechanism dynami-cally adjusts the weights of features, enabling the network to respond more flexibly to varying contextual environments. For instance, the appearance of tree trunks can differ significantly across different forest settings. By leveraging local feature attention, the network can adaptively assign greater focus to such regions, thereby enhancing model performance. This is especially beneficial in complex forest scenes, where it enables more accurate extraction of key features such as tree trunks. 2.6. Single wood splitting method For the extraction of individual tree parameters, we use the segmented point clouds as input data. After obtaining the point clouds with label information, the ground and trunk point clouds are extracted based on semantic labels, as illustrated in Fig. 6 . As shown in the Fig. 6 , the extracted point cloud effectively eliminates inter-ference from tree canopies, shrubs, vegetation, and other noise, which could otherwise affect the clustering process. Before clustering, ground normalization is required to mitigate the impact of uneven terrain on DBH estimation. This process involves gen-erating a Digital Elevation Model (DEM) with a resolution of 0.5 m using irregular tri-angular mesh interpolation (Zhao et al. 2016). The normalized point cloud is obtained by subtracting the elevation of ground points from the absolute Z-values of the data. After extracting the trunk-related point cloud data, we apply the DBSCAN algo-rithm to perform individual tree segmentation. Unlike conventional clustering algo-rithms, DBSCAN does not require prior knowledge of the number of clusters and can effectively detect arbitrarily shaped groups while filtering out noise points. It demon-strates particularly robust performance when applied to forest environments charac-terized by noise and uneven point density distributions. However, as clustering can be time-consuming on large-scale point clouds, we employed an IKd-tree to accelerate point processing and enhance overall segmentation efficiency. Upon receiving the input point cloud, the algorithm initializes all points as unvisited. Then, a point p is randomly selected from the dataset as the starting point and marked as visited. For the point p, all points within its ε-neighborhood \(\:{\text{N}}_{\text{ϵ}}\left(p\right)\) , are searched, where the neighborhood is defined by the formula in Eq. ( 1 ). $$\:\:{N}_{ϵ}\left(p\right)=\left\{q\in\:D\left|dist\right(p,q)\le\:ϵ\right\}$$ 1 In this context, D represents the dataset, and \(\:dist(p,q)\) denotes the distance between points \(\:p\) and \(\:q\) 。If the number of points \(\:{\text{N}}_{\text{ϵ}}\left(p\right)\) within the ε-neighborhood of point p satisfies \(\:{\text{N}}_{\text{ϵ}}\left(p\right)\ge\:\left|MinPts\right|\) ,then point p is labeled as a core point, and a new cluster is created by including \(\:p\) and all points in \(\:{\text{N}}_{\text{ϵ}}\left(p\right)\) ; otherwise, point p is marked as noise.After completing the above operation, the algorithm selects the next unvisited point and repeats the process until all points have been visited. For DBH estimation, this study employs the least squares circle fitting method. This approach calculates the optimal circle center and radius by minimizing the sum of the squared distances from all data points to the circumference, as shown in equations ( 2 ) and ( 3 ). Tree height is estimated using ground-referenced point clouds by performing circle fitting on the point cloud at 1.3 meters above the ground, which closely corresponds to the actual measurement height. $$\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:\:f\left({x}_{a},{y}_{a},R\right)=\sum\:{d}_{i}^{2}$$ 2 $$\:{r}_{i}=\sqrt{{({x}_{i}-{x}_{a})}^{2}+{({y}_{i}-{y}_{a})}^{2}}$$ 3 \(\:di\) represents the distance from each point at 1.3 m height to the center of the fitted circle, defined as \(\:di=({r}_{i}-R)\) . xa and ya are the coordinates of the determined center of the fitted circle, while \(\:{x}_{i}\) ​ and \(\:{y}_{i}\) are the coordinates of the circle center during the iterative fitting process. \(\:{r}_{i}\) denotes the radius of the circle fitted to each individual point, and R is the radius of the final determined fitted circle. 3. Experiment To evaluate the effectiveness of the added and improved modules in the overall system, we designed corresponding experiments and conducted quantitative assessments from four aspects: semantic segmentation accuracy (Section 3.1), system performance (Section 3.2), 3D map reconstruction (Section 3.3), and individual tree seg-mentation accuracy (Section 3.4). 3.1. Semantic segmentation accuracy To verify the accuracy of the semantic segmentation algorithm used in the system, we conducted evaluations on both a public dataset and a self-collected dataset. In ad-dition, several classical semantic segmentation algorithms were trained on the same datasets to serve as baselines for comparing and validating the accuracy of our pro-posed method. For classification results, we evaluate accuracy using average precision (Avg) and mean Intersection over Union (mIoU), and assess the real-time performance of the algorithms by measuring the number of point cloud scans processed per second (Scans/sec). 3.1.1Experimental details The experiments were conducted on a system running Ubuntu 18.04, equipped with an NVIDIA GXT4090 GPU and 24 GB of RAM, using Python 3.6 as the program-ming environment. After receiving the LiDAR-scanned point cloud, the system per-forms spherical projection on all points to convert the 3D point cloud into a 2D range image, which is then fed into the network for semantic segmentation. The 2D range image is first processed through the Efficient Local Attention (ELA) module, enabling the network to focus on key local regions, such as the junctions between tree trunks and foliage or between trunks and the ground. Subsequently, the image is passed through the semantic segmentation component of the SqueezeSegV3 algorithm, which consists of layers with output channel sizes of 64, 128, 256, 256, and 256, respectively. After segmentation, a 2D prediction label map is generated and then projected back into the 3D space. All points from the LiDAR scan are projected onto a 2D range image with a resolution of 64 × 2048. PolarNet(Zhang et al. 2020) ,SalsaNet(Aksoy et al. 2020) ,Rangnet++(Milioto et al. 2019) included in the evaluation are config-ured using their default configuration files. 3.1.2. Accuracy evaluation experiment The public dataset used in this study is the Plot_a dataset (Kaijaluoto et al. 2022), which covers the study area of Ämari, Finland (latitude 61.19°, longitude 25.11°). LiDAR data were col-lected from three test sites(A, B, and C)which were used as the training, testing, and prediction sets, respectively. The forest plot size is 32 × 32 meters. We utilized three classes from the dataset for point cloud classification: trunk, foliage, and ground. Points belonging to other classes were excluded from the analysis. Table 1 Evaluation of algorithm accuracy in the Plot_a dataset mIoU Avg Scans/sec PolarNet 0.464 0.618 13 SalsaNet 0.554 0.828 17 Rangnet++ 0.481 0.851 12 SqueezeSegV3 0.506 0.832 20 ours 0.558 0.853 18 As shown in Table 1 , on the public dataset, our algorithm performs well on both mIoU and Avg metrics. The mIoU increases by 5% compared to the algorithm before improvement, indicating higher overall segmentation quality. The Avg is also slightly higher than other algorithms, showing that it can more accurately distinguish different categories of point clouds. In addition, while maintaining improved accuracy, our al-gorithm's point cloud processing speed is slightly lower than SqueezeSegV3, but better than other algorithms, demonstrating good real-time performance. To validate the effectiveness of the algorithm, we also collected point cloud data using a lidar-equipped unmanned vehicle in two selected 25×25 m birch forest plots within the experimental area, which were used as the training set and the prediction set, respectively. We trained the model using the training set and validated it on the prediction set. During the annotation of the training point clouds, to facilitate subse-quent individual tree segmentation, the point clouds were classified into ground, trunk, foliage and others. Points belonging to other categories were set as ignore points dur-ing training. Table 2 Comparison of Algorithm Accuracy on the Self-Collected Dataset ground foliage trunk mIoU Avg Scans/sec PolarNet 0.684 0.638 0.592 0.537 0.807 14 SalsaNet 0.556 0.254 0.781 0.664 0.848 22 Rangnet++ 0.684 0.210 0.771 0.655 0.876 17 SqueezeSegV3 0.378 0.629 0.892 0.633 0.858 27 Ours 0.713 0.719 0.891 0.672 0.896 25 As shown in Table 2 , different semantic segmentation algorithms exhibit signifi-cant performance differences in forest environments. In terms of results, our algorithm performs slightly better overall than other models, demonstrating its strong perfor-mance in complex natural environments. Specifically, the confidence scores of our al-gorithm for the classes ground, foliage, and trunk are 0.713, 0.719, and 0.891, respec-tively. The overall mIoU reaches 0.772, and the average precision reaches 0.896, both significantly higher than those of RangeNet++, SalsaNet, PolarNet, and the original SqueezeSegV3 before improvement. This indicates that the introduced ELA attention mechanism enhances the model’s ability to perceive local spatial structures and con-textual semantics. By guiding the network to focus on key local regions in the forest point cloud, it significantly improves the segmentation accuracy of fine-grained targets. Although the point cloud processing efficiency of our algorithm is slightly lower than that of the original algorithm, it still exceeds the LiDAR acquisition rate (10 Hz), meet-ing the real-time processing requirements for point cloud data. Considering both ac-curacy and real-time performance in forest environments, the proposed algorithm is capable of handling the complexity of forest scenes effectively. 3.2. System Performance Evaluation Precise localization accuracy is critical for improving the quality of map recon-struction. Therefore, we collected trajectory data at the experimental site and used RTK-recorded trajectories as ground truth. We evaluated the trajectory accuracy of both our improved algorithm and the baseline method. As shown in Table Forest_01, we used the EVO tool to compute the Root Mean Square Error (RMSE), standard devi-ation (Std), and maximum error (Max) of the trajectories. Due to the complexity of forest environments, terrain conditions can hinder the unmanned vehicle from performing loop closure during data collection. To address this scenario, we recorded a long-distance dataset without loop closures, named For-est_02, to evaluate the trajectory alignment accuracy of the system under loop-closure-absent conditions, as illustrated in Fig. 10 . Table 3 RMSE, Std, and Max Errors of Different Algorithms Forest_01 Tree_02 RMSE Std Max RMSE Std Max Lio_SAM 0.0214 0.0118 0.554 0.0291 0.0174 0.558 ours 0.0213 0.0110 0.773 0.0233 0.0102 0.709 In the Forest_01 and Tree_02 scenarios, the trajectory accuracy comparison be-tween LIO-SAM and our proposed method reveals that the RMSE values of both algo-rithms are very close. However, our method achieves lower standard deviations (0.0110 and 0.0102), indicating reduced fluctuation in trajectory error and thus im-proved stability. Although the maximum error is slightly higher at 0.77 m, the overall system demonstrates stronger stability, offering enhanced accuracy and robustness in complex forested environments with irregular tree distributions. Overall, our approach improves error consistency while maintaining accuracy, showcasing better adaptability to challenging conditions. 3.2.2. Evaluation of System Runtime Efficiency This section evaluates the runtime efficiency of different algorithms when pro-cessing a single LiDAR scan. The main evaluation metrics include average processing time (Avg rate), actual memory usage (RES), and GPU utilization (GPU). The experi-mental data were collected during real-time operation in a forest environment, as shown in Table 4 ,the LIO-SAM algorithm requires a longer time to extract trunk fea-tures from raw point clouds, with an average processing time of 27.57 ms per frame and a high GPU utilization of 71.3%. In contrast, our proposed improved algorithm demonstrates significantly better performance, reducing the average runtime to 24.43 ms and lowering GPU usage to 50.5%, while maintaining a comparable memory foot-print. The improvement can be attributed to the introduction of the iKD-tree data structure during the point cloud preprocessing stage. This structure allows efficient dynamic insertion and removal of points, thereby optimizing the data processing pipe-line, reducing redundant computation, and enhancing memory utilization. Overall, our approach significantly enhances the system's real-time performance and operational efficiency. Table 4 Efficiency of different algorithms to complete a build Avg rate RES GPU Lio_SAM 27.57ms 268MB 71.3% ours 24.43ms 286MB 50.5% 3.3. 3D Map Reconstruction Accurate classification of individual tree point clouds can significantly improve the precision of clustering algorithms. To verify this, we separately collected point cloud data of a row of trees and performed semantic segmentation using the trained model. We then conducted comparative experiments with other algorithms to evaluate the effectiveness of our method. The segmentation results are presented in Fig. 10 . From the segmentation results, our algorithm accurately classified the point clouds of tree trunks, canopies, and ground. It preserved the canopy boundaries well, with no significant omissions or misclassifications. In contrast, both RangeNet + + and Squeez-eSegV3 exhibited issues such as canopy fragmentation, confusion between trunk and ground points, and blurred boundaries. Overall, after incorporating the ELA attention mechanism, our method effectively mitigates the misclassification of boundary points, thereby improving the segmentation accuracy and enhancing the overall robustness of the system. To further evaluate the feasibility of the proposed system in large-scale forest en-vironments, we conducted real-time 3D semantic reconstruction in a birch forest. A test area of 25 m × 25 m was selected, and the equipment shown in Fig. 1 was used to perform a full loop traversal at a constant speed within the forest. To facilitate subse-quent tree counting and DBH estimation, the point cloud data were semantically seg-mented into three categories: ground, trunk, and foliage. The resulting 3D map is shown in Fig. 11 . It can be observed that the system effectively classifies the point cloud into three categories: foliage, trunk, and ground. After classification, the system removes foliage points, outliers, and other irrelevant points that could affect clustering accuracy, thereby retaining a clean 3D map composed mainly of trunk and ground points. This significantly improves the precision of single-tree segmentation and clustering. Alt-hough some misclassifications may occur for high-elevation canopy points, these do not impact the estimation of tree count or DBH. Overall, the reconstructed map demonstrates the system’s capability for accurate forest data acquisition and measurement. 3.4. Evaluation of Individual Tree Segmentation Accuracy We validated the individual tree segmentation using two datasets of different tree types collected from the experimental area. As shown on the left side of Fig. 12, the point cloud was filtered by the system’s semantic segmentation module to retain only ground and trunk points. For better visualization, each cluster is displayed in a differ-ent color. As shown in the clustering results of the point cloud in Fig. 12, the system ef-fectively performs trunk point cloud clustering, with each individual tree accurately segmented into a distinct cluster, and no erroneous clustering observed. Due to the relatively low segmentation accuracy of high-elevation canopy points, we introduced a Z-axis threshold during clustering: clusters whose points all exceed this height thresh-old are discarded. As illustrated in Fig. 12(b), filtering out these high points signifi-cantly improves clustering accuracy without affecting the estimation of tree count or DBH, thereby enhancing the overall precision of the system. To evaluate the accuracy of DBH estimation, we manually measured 20 trees along the trajectory within the experimental area. The ground-truth DBH values were obtained by measuring the circumference of each tree at 1.3 meters above the ground and converting it to diameter. We then compared the actual DBH values with the fitted DBH values obtained from our system to assess the estimation accuracy. As shown in the Table 5 , we present a subset of the results along with the average error. Table 5 Assessment of tree diameter at breast height measurements ` Truth(cm) Fitted (cm) Deviation(cm) Relative Ac 1 32.5 31.1 1.1 97.2% 2 31.1 30.5 0.6 98.0% 3 14.4 15.4 1.0 93.5% 4 23.7 24.6 0.9 96.3% 5 23.0 23.6 0.6 97.4% 6 21.2 21.0 0.2 99.1% 7 16.8 17.2 0.4 97.7% 8 15.9 15.5 0.4 97.5% 9 17.2 16.4 0.8 95.3% 10 17.5 18.0 0.8 97.2% Avg 21.2 21.5 0.58 98.6% The collected LiDAR point cloud data were used as experimental input for the system. As shown in Table 3 , the DBH fitting results demonstrate overall good per-formance, with an average deviation of 0.58 cm. Considering the average actual DBH of 21.2 cm, the error is relatively small, resulting in an average relative accuracy of 98.6%, indicating that the model has strong fitting capability. Among the 10 samples, most fitting errors were within 1 cm, with only samples No. 1 and No. 3 exceeding 1.0 cm, yet still within the acceptable range of forestry measurement standards. In terms of relative accuracy, all but one sample achieved over 95%, with most exceeding 97%. Overall, the fitting results show high accuracy and consistency, demonstrating the system’s effectiveness for forest applications. 4. Discussion In this study, we presented a lightweight LiDAR SLAM system integrated with semantic segmentation, specifically designed to meet the demands of real-time per-formance and accurate individual tree segmentation in forest environments. To enhance real-time capability, the system incorporates the IKD-Tree algorithm, which constructs a dynamic KD-tree structure to significantly improve nearest neighbor search efficiency during feature matching. This optimization accelerates point cloud registration and feature extraction after downsampling, resulting in a 21.2% reduction in CPU usage compared to the original LIO-SAM algorithm and a 3 ms decrease in processing latency. In addition, the improved SqueezeSegV3 network integrated into the system achieved an mIoU of 0.67 in forestry scenarios, representing a 4% improvement over the original model. The accuracy of DBH measurement after clustering reached 98.6%, with an av-erage deviation of 0.58 cm, indicating that the system performs reliably in tree identi-fication, clustering, and DBH estimation, with a prediction error within 3%. Overall, compared with traditional LiDAR scanning systems used in forestry applications, our approach maintains high accuracy while significantly reducing system complexity, making it more suitable for deployment on mobile platforms in forest environments. In practical applications, the integration of semantic segmentation enables the effective removal of non-target point clouds such as weeds and shrubs, thereby improving the accuracy of trunk clustering and DBH estimation. These results demonstrate the sys-tem's robustness and practicality for forest resource assessment tasks. Our research integrates a lightweight semantic segmentation network with an ef-ficient SLAM framework, providing a deployable and scalable solution for tasks such as forest resource investigation and individual tree extraction. However, the current system achieves high segmentation accuracy primarily for tree trunks beneath the canopy, while point clouds in the upper canopy layers tend to be misclassified, making accurate estimation of tree height parameters challenging. In future work, we aim to further optimize the network architecture to develop a more efficient and accurate deep learning model, and to explore effective methods for precise tree height estima-tion, thereby advancing the application of this system in autonomous forest resource surveys using unmanned vehicles. 5. Conclusion In this paper, we proposed a lightweight LiDAR SLAM system integrated with semantic segmentation and deployed it on an unmanned vehicle. The system incorpo-rates an improved lightweight semantic segmentation network, SqueezeSegV3, to en-hance the accuracy of individual tree segmentation, and employs a dynamic IKD-Tree structure to accelerate point cloud feature matching and data association. Experimental results demonstrate that the proposed method effectively addresses the challenges of low processing efficiency and limited segmentation accuracy in forest point cloud data. It achieves strong performance in tasks such as tree identification and DBH estimation, showing high practical value. In future work, we aim to extend the system's capabilities to tree height estimation and complex canopy structure extraction, promoting the broader application of autonomous vehicles in intelligent forest resource surveys and mobile operational platforms. Declarations Author Contribution Conceptualization, S.Z. and Y.Z. (Yongxu Zhou).; methodology, S.Z.; software, Y.Z.(Yunfeng Zhao); formal analysis, B.S.; investigation, Y.Z(Yunfeng Zhao).; data curation, B.S. and J.W..; writing—original draft preparation, S.Z. and Y.Z. (Yongxu Zhou); project administration, Y.Z. (Yunfeng Zhao); funding acquisition, J.W. All authors have read and agreed to the published version of the manuscript. References Aksoy EE, Baci S, Cavdar S (2020) SalsaNet: Fast Road and Vehicle Segmentation in LiDAR Point Clouds for Autonomous Driving. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp 926–932 Bárta V, Hanuš J, Dobrovolný L, Homolová L (2022) Comparison of field survey and remote sensing techniques for detection of bark beetle-infested trees. Forest Ecology and Management 506:119984. https://doi.org/10.1016/j.foreco.2021.119984 Bauwens S, Bartholomeus H, Calders K, Lejeune P (2016) Forest Inventory with Terrestrial LiDAR: A Comparison of Static and Hand-Held Mobile Laser Scanning. Forests 7:127. https://doi.org/10.3390/f7060127 Brede B, Calders K, Lau A, et al (2019) Non-destructive tree volume estimation through quantitative structure modelling: Comparing UAV laser scanning with terrestrial LIDAR. Remote Sensing of Environment 233:111355. https://doi.org/10.1016/j.rse.2019.111355 Cai Y, Xu W, Zhang F (2021) ikd-Tree: An Incremental K-D Tree for Robotic Applications Chen SW, Nardari GV, Lee ES, et al (2020) SLOAM: Semantic Lidar Odometry and Mapping for Forest Inventory. IEEE Robot Autom Lett 5:612–619. https://doi.org/10.1109/LRA.2019.2963823 Eeckhaut MVD, Poesen J, Verstraeten G, et al (2007) Use of LIDAR‐derived images for mapping old landslides under forest. Earth Surf Processes Landf 32:754–769. https://doi.org/10.1002/esp.1417 Fadil S, Sebari I, Moulay MA, kadi KAE (2024) Modeling and spatialization of biomass and carbon stock using unmanned Aerial Vehicle Lidar (Lidar-UAV) metrics and forest inventory in cork oak forest of Maamora. Regional Science Policy & Practice 16:100127. https://doi.org/10.1016/j.rspp.2024.100127 Fan Y, Feng Z, Shen C, et al (2020) A trunk-based SLAM backend for smartphones with online SLAM in large-scale forest inventories. ISPRS Journal of Photogrammetry and Remote Sensing 162:41–49. https://doi.org/10.1016/j.isprsjprs.2020.02.006 Gao Q, Kan J (2022) Automatic Forest DBH Measurement Based on Structure from Motion Photogrammetry. Remote Sensing 14:2064. https://doi.org/10.3390/rs14092064 Gollob C, Ritter T, Kraßnitzer R, et al (2021) Measurement of Forest Inventory Parameters with Apple iPad Pro and Integrated LiDAR Technology. Remote Sensing 13:3129. https://doi.org/10.3390/rs13163129 Gollob C, Ritter T, Nothdurft A (2020) Forest Inventory with Long Range and High-Speed Personal Laser Scanning (PLS) and Simultaneous Localization and Mapping (SLAM) Technology. Remote Sensing 12:1509. https://doi.org/10.3390/rs12091509 Hao Y, Widagdo FRA, Liu X, et al (2020) Individual Tree Diameter Estimation in Small-Scale Forest Inventory Using UAV Laser Scanning. Remote Sensing 13:24. https://doi.org/10.3390/rs13010024 He GB, Li LL (2020) RESEARCH AND APPLICATION OF LIDAR TECHNOLOGY IN CADASTRAL SURVEYING AND MAPPING. Int Arch Photogramm Remote Sens Spatial Inf Sci XLIII-B1-2020:33–37. https://doi.org/10.5194/isprs-archives-XLIII-B1-2020-33-2020 Kaijaluoto R, Kukko A, El Issaoui A, et al (2022) Semantic segmentation of point cloud data using raw laser scanner measurements and deep neural networks. ISPRS Open Journal of Photogrammetry and Remote Sensing 3:100011. https://doi.org/10.1016/j.ophoto.2021.100011 Kukko A, Kaijaluoto R, Kaartinen H, et al (2017) Graph SLAM correction for single scanner MLS forest data under boreal forest canopy. ISPRS Journal of Photogrammetry and Remote Sensing 132:199–209. https://doi.org/10.1016/j.isprsjprs.2017.09.006 Li J, Liu J, Huang Q (2023) PointDMM: A Deep-Learning-Based Semantic Segmentation Method for Point Clouds in Complex Forest Environments. Forests 14:2276. https://doi.org/10.3390/f14122276 Liu B, Huang H, Su Y, et al (2022) Tree Species Classification Using Ground-Based LiDAR Data by Various Point Cloud Deep Learning Methods. Remote Sensing 14:5733. https://doi.org/10.3390/rs14225733 Ma Z, Dong Y, Zi J, et al (2023) Forest-PointNet: A Deep Learning Model for Vertical Structure Segmentation in Complex Forest Scenes. Remote Sensing 15:4793. https://doi.org/10.3390/rs15194793 Milioto A, Vizzo I, Behley J, Stachniss C (2019) RangeNet ++: Fast and Accurate LiDAR Semantic Segmentation. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp 4213–4220 Mu X, Zhao D, Zheng Z, et al (2025) Can Stereoscopic Density Replace Planar Density for Forest Aboveground Biomass Estimation? A Case Study Using Airborne LiDAR and Landsat Data in Daxing’anling, China. Remote Sensing 17:1163. https://doi.org/10.3390/rs17071163 Neuville R, Bates JS, Jonard F (2021) Estimating Forest Structure from UAV-Mounted LiDAR Point Cloud Using Machine Learning. Remote Sensing 13:352. https://doi.org/10.3390/rs13030352 Nuchter A, Lingemann K, Hertzberg J (2007) Cached k-d tree search for ICP algorithms. In: Sixth International Conference on 3-D Digital Imaging and Modeling (3DIM 2007). pp 419–426 Oveland I, Hauglin M, Giannetti F, et al (2018) Comparing Three Different Ground Based Laser Scanning Methods for Tree Stem Detection. Remote Sensing 10:538. https://doi.org/10.3390/rs10040538 Pierzchała M, Giguère P, Astrup R (2018) Mapping forests using an unmanned ground vehicle with 3D LiDAR and graph-SLAM. Computers and Electronics in Agriculture 145:217–225. https://doi.org/10.1016/j.compag.2017.12.034 Qian C, Liu H, Tang J, et al (2016) An Integrated GNSS/INS/LiDAR-SLAM Positioning Method for Highly Accurate Forest Stem Mapping. Remote Sensing 9:3. https://doi.org/10.3390/rs9010003 Rajab Pourrahmati M, Baghdadi N, Fayad I (2023) Comparison of GEDI LiDAR Data Capability for Forest Canopy Height Estimation over Broadleaf and Needleleaf Forests. Remote Sensing 15:1522. https://doi.org/10.3390/rs15061522 Shan T, Englot B, Meyers D, et al (2020) LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping Tang J, Chen Y, Kukko A, et al (2015) SLAM-Aided Stem Mapping for Forest Inventory with Small-Footprint Mobile LiDAR Tian L, Wu X, Tao Y, et al (2023) Review of Remote Sensing-Based Methods for Forest Aboveground Biomass Estimation: Progress, Challenges, and Prospects. Forests 14:1086. https://doi.org/10.3390/f14061086 Wang Y, Lin Y, Cai H, Li S (2022) Hierarchical Fine Extraction Method of Street Tree Information from Mobile LiDAR Point Cloud Data. Applied Sciences 13:276. https://doi.org/10.3390/app13010276 Xu C, Wu B, Wang Z, et al (2021a) SqueezeSegV3: Spatially-Adaptive Convolution for Efficient Point-Cloud Segmentation Xu D, Wang H, Xu W, et al (2021b) LiDAR Applications to Estimate Forest Biomass at Individual Tree Scale: Opportunities, Challenges and Future Perspectives. Forests 12:550. https://doi.org/10.3390/f12050550 Xu W, Wan Y (2024) ELA: Efficient Local Attention for Deep Convolutional Neural Networks Yang S, Xing Y, Xing T, et al (2024) Multisensors Fusion SLAM-Aided Forest Plot Mapping With Backpack Dual-LiDAR System. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 17:16051–16070. https://doi.org/10.1109/JSTARS.2024.3451175 Yu J, Lei L, Li Z (2024) Individual Tree Segmentation Based on Seed Points Detected by an Adaptive Crown Shaped Algorithm Using UAV-LiDAR Data. Remote Sensing 16:825. https://doi.org/10.3390/rs16050825 Zhang T, Long J, Lin H, et al (2024) A Novel Feature Evaluation Method in Mapping Forest AGB by Fusing Multiple Evaluation Metrics Using PolSAR Data. IEEE Geoscience and Remote Sensing Letters 21:1–5. https://doi.org/10.1109/LGRS.2024.3378425 Zhang Y, Zhou Z, David P, et al (2020) PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, Seattle, WA, USA, pp 9598–9607 Zhao X, Guo Q, Su Y, Xue B (2016) Improved progressive TIN densification filtering algorithm for airborne LiDAR data in forested areas. ISPRS Journal of Photogrammetry and Remote Sensing 117:79–91. https://doi.org/10.1016/j.isprsjprs.2016.03.016 Zhou T, Zhao C, Wingren CP, et al (2024) Forest feature LiDAR SLAM (F2-LSLAM) for backpack systems. ISPRS Journal of Photogrammetry and Remote Sensing 212:96–121. https://doi.org/10.1016/j.isprsjprs.2024.04.025 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 11 Sep, 2025 Reviews received at journal 08 Sep, 2025 Reviews received at journal 02 Sep, 2025 Reviewers agreed at journal 27 Aug, 2025 Reviewers agreed at journal 26 Aug, 2025 Reviews received at journal 15 Aug, 2025 Reviewers agreed at journal 09 Aug, 2025 Reviewers invited by journal 08 Jul, 2025 Editor assigned by journal 03 Jul, 2025 Submission checks completed at journal 01 Jul, 2025 First submitted to journal 25 Jun, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6971887","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":482498053,"identity":"9f4fb3cf-cc7c-46ee-a6c0-290e049346dd","order_by":0,"name":"Yunfeng Zhao","email":"","orcid":"","institution":"North China Institute of Aerospace Engineering","correspondingAuthor":false,"prefix":"","firstName":"Yunfeng","middleName":"","lastName":"Zhao","suffix":""},{"id":482498055,"identity":"360d9324-ac58-4991-a4d6-40459bd26c65","order_by":1,"name":"Shipeng Zhao","email":"","orcid":"","institution":"North China Institute of Aerospace Engineering","correspondingAuthor":false,"prefix":"","firstName":"Shipeng","middleName":"","lastName":"Zhao","suffix":""},{"id":482498056,"identity":"da703fcb-b7a9-4a72-b231-8ea23de6aa28","order_by":2,"name":"Yongxu Zhou","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA3klEQVRIiWNgGAWjYBACAwkeMC3HwMDYAGJASAJawIqMSdeSCFNJWIu5dO/xhz/bDqdvZz/c+OEDg43shgPMzx7g02I551xig2RbWu7OnsRmyRkMacYbDrCZG+B12I0cwwbDNpvcDTcYG6R5GA4nbjjAwyZBUEtim0S6wQ3G5t9/GP4TqeVgm00CUEubNAPDAcJaLGfkGM5sOJdmuOFMYptlj0Gy8czDbGZ4tZhL5Bh8/FF2WN7g+PHHN35U2Mn2HW9+hlcLGDCywd0JxMwE1YPAH6JUjYJRMApGwUgFAM99TY6ky6/YAAAAAElFTkSuQmCC","orcid":"","institution":"North China Institute of Aerospace Engineering","correspondingAuthor":true,"prefix":"","firstName":"Yongxu","middleName":"","lastName":"Zhou","suffix":""},{"id":482498058,"identity":"5844f788-b5e9-476b-83c1-7081c84501d8","order_by":3,"name":"Bing Shen","email":"","orcid":"","institution":"Beijing Institute of Tracking and Telecommunication Technology","correspondingAuthor":false,"prefix":"","firstName":"Bing","middleName":"","lastName":"Shen","suffix":""},{"id":482498059,"identity":"a6d9daf6-2443-4910-afb8-3d1c85e195d1","order_by":4,"name":"Jing Wang","email":"","orcid":"","institution":"Beijing Institute of Tracking and Telecommunication Technology","correspondingAuthor":false,"prefix":"","firstName":"Jing","middleName":"","lastName":"Wang","suffix":""}],"badges":[],"createdAt":"2025-06-25 07:38:38","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6971887/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6971887/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":86409560,"identity":"4ed71871-70ec-41d3-84ee-c27ac9caf266","added_by":"auto","created_at":"2025-07-10 10:30:44","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":395960,"visible":true,"origin":"","legend":"\u003cp\u003eForest Inventory Experimental Platform\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/233a644567ed38d53891ff30.png"},{"id":86409555,"identity":"5b706690-7ad3-48d9-9438-dbc338d4bb9d","added_by":"auto","created_at":"2025-07-10 10:30:44","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":293714,"visible":true,"origin":"","legend":"\u003cp\u003eStudy area(a) Collection area selection (b) Field data collection site (c) Point cloud processing tools\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/40289c5010553d43b4808f8a.png"},{"id":86411200,"identity":"a87a0e4f-e2de-4113-bcc9-5ed32744bd45","added_by":"auto","created_at":"2025-07-10 10:38:44","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":269005,"visible":true,"origin":"","legend":"\u003cp\u003eSystem Technical Workflow.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/a765e307538c6b51ffcf10a9.png"},{"id":86409556,"identity":"e71cf3a9-dfbc-4525-a259-10b2189c836c","added_by":"auto","created_at":"2025-07-10 10:30:44","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":152531,"visible":true,"origin":"","legend":"\u003cp\u003eRe-build an unbalanced sub-tree\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/65c771f464f9f1f609b125b5.png"},{"id":86409559,"identity":"859bde70-934a-419d-9b79-ee098f842474","added_by":"auto","created_at":"2025-07-10 10:30:44","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":276741,"visible":true,"origin":"","legend":"\u003cp\u003eSemantic Segmentation Algorithm Process Framework\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/af3dedbb27fe4b2b573bc929.png"},{"id":86409578,"identity":"a3f87d09-5547-4470-968e-ac54ef41f888","added_by":"auto","created_at":"2025-07-10 10:30:44","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":295030,"visible":true,"origin":"","legend":"\u003cp\u003eSchematic of the extracted point cloud.(a) Point cloud with semantic la-bels; (b) Extracted ground and tree trunk points\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/7fc61e0ef860a66afbbdb3f9.png"},{"id":86412106,"identity":"079de828-7c3b-4dfc-a40a-037be8fa8cdd","added_by":"auto","created_at":"2025-07-10 10:46:44","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":87116,"visible":true,"origin":"","legend":"\u003cp\u003eLinear fitting of point cloud for DBH\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/fce0450ec649592501720d59.png"},{"id":86409563,"identity":"c984954d-589b-4d96-8e47-ab31af7846d5","added_by":"auto","created_at":"2025-07-10 10:30:44","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":91550,"visible":true,"origin":"","legend":"\u003cp\u003e(a) Top: 2D projection of point cloud; (a) Bottom: segmentation result of the point cloud after 2D projection; (b) Comparison between the original point cloud and the segmentation result after re-projection.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/662b8411c5801d4cca509a7b.png"},{"id":86409569,"identity":"83617bea-4cc0-43f3-819d-aa09ec5e6b72","added_by":"auto","created_at":"2025-07-10 10:30:44","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":166361,"visible":true,"origin":"","legend":"\u003cp\u003eTrajectory Diagram of Forest_01.(a) Space trajectory schematic, (b) Plane trajectory schematic\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/5beb96dcd2954bc7f4632101.png"},{"id":86411202,"identity":"ac93a65b-5e08-4e60-98b1-1c3a57ff87ab","added_by":"auto","created_at":"2025-07-10 10:38:44","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":156637,"visible":true,"origin":"","legend":"\u003cp\u003eTrajectory Diagram of Forest_02. (a) Space trajectory schematic, (b) Plane trajectory schematic\u003c/p\u003e","description":"","filename":"10.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/a95c66115e94f4ac8d198989.png"},{"id":86411258,"identity":"369e23bf-e825-4f2b-b18d-b7c8b4374ff2","added_by":"auto","created_at":"2025-07-10 10:38:45","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":522336,"visible":true,"origin":"","legend":"\u003cp\u003eFig. 10. Comparison of semantic segmentation results on point clouds: (a) Original point cloud intensity map; (b) Improved V3 algorithm; (c) RangeNet++; (d) Squeez-eSegV3.\u003c/p\u003e","description":"","filename":"11.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/266ae0eb0cc391f9c4b9c727.png"},{"id":86411254,"identity":"38a5718b-6d96-4a94-931c-10d3946dd589","added_by":"auto","created_at":"2025-07-10 10:38:45","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":608227,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFig. 11\u003c/strong\u003e. Forest point cloud 3D reconstruction map\u003c/p\u003e","description":"","filename":"12.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/95d037988e88ef879c1157b2.png"},{"id":86409586,"identity":"5aba8903-3ae1-431b-8278-1b6628096692","added_by":"auto","created_at":"2025-07-10 10:30:45","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":269331,"visible":true,"origin":"","legend":"\u003cp\u003eFig. 11. Point cloud clustering results. (a) Self-built dataset (b) Public dataset\u003c/p\u003e","description":"","filename":"13.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/a2da0d38a8d6f734c49a1b3c.png"},{"id":86409573,"identity":"c7ea6e3e-ea4a-4e58-8e2e-dc785480a0b5","added_by":"auto","created_at":"2025-07-10 10:30:44","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":526096,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFig.13.\u003c/strong\u003e Tree diameter measurement and data acquisition area. (a) Manual data collection method; (b) Data collection route of the unmanned vehicle. area\u003c/p\u003e","description":"","filename":"14.png","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/d2e0ad8067dd284136caf35c.png"},{"id":86413371,"identity":"260ea1a4-33a5-4dfe-9966-7760748feeca","added_by":"auto","created_at":"2025-07-10 10:54:47","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":5147958,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6971887/v1/2988335b-b879-4d5b-84bd-fb2b3e3f5881.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"A Lightweight LiDAR SLAM System Integrated with Semantic Segmentation for Forest Biomass Parameter Acquisition","fulltext":[{"header":"Key message ","content":"\u003cp\u003eA lightweight SLAM system integrated with semantic segmentation is proposed, enabling efficient and accurate single-tree extraction and diameter at breast height (DBH) measurement in complex forest environments.\u0026nbsp;\u003c/p\u003e"},{"header":"1. Introduction","content":"\u003cp\u003eAboveground forest biomass(tree height, DBH, and species composition)is a fundamental component and key indicator in forest inventory and assessment (Zhang et al. 2024,Mu et al. 2025). Conventional forest resource surveys are typically conducted through manual measurements, which demand substantial labor and time and may also negatively impact the natural environment(Gollob et al. 2021). Advances in sensor technologies have facilitated the wide-spread use of remote and proximal sensing techniques in automated forest surveys across multiple spatial scales(Tian et al. 2023,B\u0026aacute;rta et al. 2022). However, due to its inability to capture the three-dimensional structure of forests, optical imagery is limited in providing com-prehensive biomass estimations. The accurate acquisition of forest structural details and key parameters is currently achieved primarily through Terrestrial Laser Scanning (TLS), Unmanned Aerial Vehicle Laser Scanning (UAV-LS), and Mobile Laser Scanning (MLS) (Xu et al. 2021b). Terrestrial laser scanning is time-consuming in data acquisition (Brede et al. 2019) and is ineffective in estimating individual tree parameters under occluded conditions(Bauwens et al. 2016), making it un-suitable for complex forest environments. UAV-LS, which is not constrained by forest terrain, offers advantages such as rapid data acquisition and ease of deployment, making it widely used for forest data collection. Neuville et al. (2021) utilized UAV-acquired LiDAR data and applied machine learning techniques to estimate forest structure. Similarly, Fadi et al. (2024) used UAV data combined with allometric equations to estimate forest biomass and carbon storage. Previous studies have demonstrated the effective-ness of UAV in forest data collection. However, since UAVs primarily operate above the canopy, foliage occlusion often prevents accurate measurements of sub-canopy forest structure and ground-level forest biometric parameters (Hao et al. 2020). In comparison, MLS sys-tems excel in ground-based coverage, enabling detailed 3D reconstruction of forest understory architecture and significantly improving single-tree feature extraction ac-curacy (Gao and Kan 2022).\u003c/p\u003e\u003cp\u003eMobile Laser Scanning is mainly divided into airborne and backpack-mounted systems. By integrating Simultaneous Localization and Mapping (SLAM) technology, MLS can effectively acquire large-scale forest data (Kukko et al. 2017). SLAM relying on its excellent real-time performance and high-precision mapping capabilities, has been widely ap-plied in forestry resource surveys (Fan et al. 2020,Gollob et al. 2020), Tang et al. (2015)investigated a LiDAR-IMU SLAM system mounted on an all-terrain vehicle (ATV) for forest inventory applications. Building upon this work, Qian et al. (2016) incorporated GNSS measurements into the system to achieve higher positioning accuracy. Pierzchała et al. (2018) utilized 3D LiDAR combined with SLAM to generate forest maps and estimate individual tree positions and diameters at breast height (DBH). Besides airborne LiDAR acquisition, back-pack-mounted laser scanning has also been widely employed. For example, Oveland et al. (2018) developed a backpack system equipped with two 3D LiDAR sensors featuring vertical and horizontal rotation axes. This dual-LiDAR configuration enables the system to cover a wider area during scanning and capture more feature information from multiple angles. Zhou et al. (2024) proposed a backpack-mounted laser scanning system with the F2-SLAM algorithm, which extracts ground and tree trunk features from LiDAR data and employs least squares adjustment to optimize LiDAR scan registration. The system also integrates trajectory from the navigation unit to enhance feature extraction. The above studies demonstrate that integrating SLAM technology with LiDAR for forest data collection enables fast, accurate, and comprehensive acquisition of the three-dimensional structure of trees and vegetation (Yang et al. 2024), thereby improving data col-lection efficiency and reducing the time and cost associated with manual measure-ments.\u003c/p\u003e\u003cp\u003eHowever, due to the complex spatial structure of forests, relying solely on 3D LiDAR-based SLAM technology provides limited geometric information (Wang et al. 2022), making it difficult to handle scenarios with highly similar tree structures and uneven density distributions. For instance, densely clustered multi-stem trees may be mistakenly identified as a single treetop (Yu et al. 2024), resulting in poor segmentation accuracy (Rajab Pourrahmati et al. 2023). Addi-tionally, large treetop regions may be incorrectly detected as individual trees, leading to segmentation errors. These issues significantly affect the reliability of single-tree ex-traction. To address this challenge and improve segmentation accuracy, this study in-corporates deep learning to obtain semantic information of the forest, thereby providing the system with more robust feature representations. Deep learning has become an ef-fective solution for addressing the challenges of single-tree segmentation in complex forest environments. For example, Chen et al. (2020) proposed SLOAM, a semantic LiDAR SLAM system for forest resource inventory that can extract individual tree information through instance segmentation and automatically estimate DBH. However, this system is prone to interference from underbrush noise during ground point cloud segmentation. Liu et al. (2022) presented a LiDAR point cloud semantic segmentation method tailored for complex forest environments, capable of accurately identifying underbrush and individual trees, but the algorithm lacks spatial hierarchical awareness of forest structure, making it difficult to adapt to complex scenarios. Ma et al. (2023) introduced the Forest-PointNet model specifically designed to extract vertical forest structural information from terrestrial LiDAR data, achieving high-precision semantic segmentation of different structural layers in complex forest scenes. Li et al. (2023) pro-posed the Point DMM method, which combines structured annotation with multi-scale feature extraction to effectively improve segmentation accuracy and robustness. How-ever, its complex model architecture entails a high demand for computational re-sources. The aforementioned studies demonstrate the effectiveness and feasibility of deep learning techniques in acquiring tree parameters in forestry applications. Never-theless, achieving high-precision segmentation typically relies on large-scale point cloud datasets, which poses challenges to system efficiency and real-time performance, especially when deploying on embedded platforms such as unmanned vehicles. Therefore, developing a lightweight and real-time semantic SLAM system has become a current research focus. Such a system not only facilitates fine-grained classification of forest resources but also provides efficient and reliable data support for individual tree parameter extraction, stand structure analysis, and precise biomass estimation.\u003c/p\u003e\u003cp\u003eTo address the aforementioned issues, we propose a LiDAR SLAM-based forestry tree resource inventory system integrated with deep learning. The system primarily utilizes unmanned vehicles and LiDAR sensors to generate semantic point cloud maps of forests, enabling the measurement of DBH and tree count. To achieve more accurate and real-time output of forest internal structure information, we employ the V3 algo-rithm to process LiDAR point cloud data, extracting features such as trunks and ground through semantic segmentation. Compared with traditional point cloud clustering methods, such as LiDAR odometry and mapping (He and Li 2020),(Eeckhaut et al. 2007), the proposed system demonstrates superior accuracy and efficiency. The main contributions of this study include:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003epropose a lightweight semantic LiDAR SLAM system that integrates an improved LIO-SAM algorithm with the semantic segmentation network SqueezeSegV3 to construct real-time 3D maps. By fusing semantic information with point clouds, the system enriches the point cloud with semantic features and enhances mapping accuracy.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eincorporate the incremental k-d tree (IKD-Tree) algorithm for dynamic point cloud updates, which significantly improves the processing speed of the SLAM system and ensures real-time performance in large-scale environments.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eintroduce a semantic segmentation-based individual tree extraction method, which reduces the influence of noise on point cloud clustering and improves the accuracy of individual tree identification.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e"},{"header":"2. Material and methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1. Hardware equipment\u003c/h2\u003e\u003cp\u003eThe experimental platform used in this study for forest inventory is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. The unmanned vehicle chassis measures 0.726 m in length, 0.617 m in width, and 0.273 m in height, with a maximum payload capacity of 50 kg. It adopts differential steering and has a maximum operating speed of 0.8 m/s, with a climbing ability of up to 25\u0026deg;. The vehicle supports manual control via a joystick, with a maximum control range of 20 meters.The vehicle is equipped with a MID360 LiDAR sensor, which collects 20,000 point cloud data points per second at a frequency of 10 Hz. The sensor features a 360\u0026deg; horizontal and a vertical viewing range from \u0026minus;\u0026thinsp;7\u0026deg; to 52\u0026deg;. It is managed by an integrated PC acting as the upper computer, powered by an AMD Ryzen3 3200G CPU (3.6 GHz). The system operates on Ubuntu 18.04 with the Robot Operating System (ROS) frame-work.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2. Study area\u003c/h2\u003e\u003cp\u003eThe study area selected for this research is Lihu Forest Park, located in Gu'an County, Langfang City, Hebei Province. The park consists of a mix of natural forests and artificially planted woodlands. Two forest plots with different tree species within this area were selected as data collection sites. Each plot covers an area of 25m*25m. The collected data from each plot were partitioned and used for training and testing purposes.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3. System Overview\u003c/h2\u003e\u003cp\u003eIn This study adopts LIO-SAM (Shan et al. 2020) as the foundational algorithm and makes im-provements tailored to the application scenario, as illustrated in the system framework shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. The system comprises three main components: a SLAM module, a semantic segmentation module, and an individual tree extraction module. After the LiDAR mounted on the unmanned vehicle acquires point cloud data, the data are first input into both the SLAM and semantic segmentation modules. In the SLAM module, to enhance frontend processing efficiency, the original downsampling algorithm is re-placed with the ikd-tree method (Cai et al. 2021), accelerating feature extraction and reducing computational load. The semantic segmentation module utilizes an improved Squeez-eSegV3 algorithm (Xu et al. 2021a) combined with a local attention mechanism, ELA (Xu and Wan 2024), effec-tively enhancing the recognition of features such as ground and tree trunks. The seg-mented point clouds are then fed back to the SLAM module to improve mapping ac-curacy. Meanwhile, the individual tree extraction module uses the segmented point clouds as the basis for subsequent processing. During individual tree extraction, the system obtains classified trunk and ground point clouds, applies the DBSCAN clus-tering algorithm to identify individual trees, and performs cylindrical fitting through linear fitting methods to estimate tree DBH.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.4. Ikd-tree algorithm\u003c/h2\u003e\u003cp\u003eThe K-D Tree is an efficient data structure that organizes multidimensional point data, enabling fast nearest neighbor searches (Nuchter et al. 2007). However, the K-D Tree is inherently static, requiring reconstruction and merging from scratch each time a new data frame is received. This process is time-consuming and negatively impacts system efficiency, making it unsuitable for lightweight and real-time applications. Therefore, we adopt the more efficient IKD-Tree algorithm to replace the original K-D Tree in our system.\u003c/p\u003e\u003cp\u003eThe IKd-Tree is a dynamic data structure based on the K-D tree, with its core composed of a binary search tree capable of incremental updates. Each node corresponds to the minimum bounding rectangle defined by its associated point set and stores the splitting axis along with the corresponding threshold to divide the node into two subregions. To support real-time performance, the IKd-Tree introduces incremental operations such as point insertion, reinsertion, and deletion. These operations are performed through recursive updates of local node states, and when necessary, local subtree reconstruction is triggered to maintain overall structural balance and query efficiency. In addition, the algorithm incorporates a lightweight dynamic rebalancing mechanism to prevent efficiency degradation caused by structural imbalance during construction and long-term operation. Specifically, when a subtree imbalance is detected, parts of the structure are selectively decomposed and locally reconstructed in parallel. This approach preserves accuracy while minimizing the impact of rebalancing on the system\u0026rsquo;s overall runtime performance. By leveraging localized reconstruction and multithreaded parallelism, the IKd-Tree maintains high responsiveness and efficient query performance even under high-frequency updates and large-scale point cloud processing scenarios.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIn the IKd-Tree, point insertion is performed by sequentially comparing coordi-nate values along the splitting axes to determine the appropriate position within the tree. Based on the local structural state of the tree, subtrees may be selectively recon-structed to maintain overall balance. Theoretically, in the worst case, an insertion op-eration may require up to \u0026#119874;(\u0026#119899;)comparisons. However, thanks to dynamic maintenance strategies and point count control mechanisms, the actual insertion efficiency is typi-cally stabilized at logarithmic complexity in practice. For point deletion, the IKd-Tree similarly employs a recursive search to locate the target node, followed by local sub-tree reconstruction when necessary to maintain structural balance. During query oper-ations, the algorithm evaluates each level based on the splitting dimension and recur-sively traverses the left and right subtrees. In most cases, the query complexity re-mains at O(log n). By incorporating rebalancing mechanisms during both insertion and deletion stages, the IKd-Tree significantly enhances computational efficiency and structural stability in large-scale point cloud processing tasks.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e2.5. semantic segmentation framework\u003c/h2\u003e\u003cp\u003eSemantic segmentation helps to compensate for the lack of feature information in LiDAR data under complex environments. In this study, we adopt the SqueezeSegV3 algorithm as the core semantic segmentation framework. This algorithm is a light-weight segmentation network based on range images and leverages spherical projec-tion to map 3D point clouds onto 2D spherical grids, resulting in dense range imag-es.To better handle the spatially varying feature distributions of LiDAR images, the algorithm introduces a Spatially-Adaptive Convolution (SAC) module. This module provides efficient spatial adaptability and content-awareness, making it well-suited for forest environments with complex spatial structures. The semantic segmentation workflow is illustrated in the Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo improve the performance of the segmentation algorithm in forest environ-ments, we integrate the Efficient Local Attention (ELA) mechanism into our frame-work. After obtaining the 2D range image, the image features are first passed through the ELA module, which accurately captures the location of regions of interest. This method maintains the original channel dimensions of the input feature maps and pre-serves the lightweight nature of the network.The local attention mechanism dynami-cally adjusts the weights of features, enabling the network to respond more flexibly to varying contextual environments. For instance, the appearance of tree trunks can differ significantly across different forest settings. By leveraging local feature attention, the network can adaptively assign greater focus to such regions, thereby enhancing model performance. This is especially beneficial in complex forest scenes, where it enables more accurate extraction of key features such as tree trunks.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003e2.6. Single wood splitting method\u003c/h2\u003e\u003cp\u003eFor the extraction of individual tree parameters, we use the segmented point clouds as input data. After obtaining the point clouds with label information, the ground and trunk point clouds are extracted based on semantic labels, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAs shown in the Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, the extracted point cloud effectively eliminates inter-ference from tree canopies, shrubs, vegetation, and other noise, which could otherwise affect the clustering process. Before clustering, ground normalization is required to mitigate the impact of uneven terrain on DBH estimation. This process involves gen-erating a Digital Elevation Model (DEM) with a resolution of 0.5 m using irregular tri-angular mesh interpolation (Zhao et al. 2016). The normalized point cloud is obtained by subtracting the elevation of ground points from the absolute Z-values of the data.\u003c/p\u003e\u003cp\u003eAfter extracting the trunk-related point cloud data, we apply the DBSCAN algo-rithm to perform individual tree segmentation. Unlike conventional clustering algo-rithms, DBSCAN does not require prior knowledge of the number of clusters and can effectively detect arbitrarily shaped groups while filtering out noise points. It demon-strates particularly robust performance when applied to forest environments charac-terized by noise and uneven point density distributions. However, as clustering can be time-consuming on large-scale point clouds, we employed an IKd-tree to accelerate point processing and enhance overall segmentation efficiency.\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eUpon receiving the input point cloud, the algorithm initializes all points as unvisited. Then, a point p is randomly selected from the dataset as the starting point and marked as visited. For the point p, all points within its ε-neighborhood \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{N}}_{\\text{ϵ}}\\left(p\\right)\\)\u003c/span\u003e\u003c/span\u003e, are searched, where the neighborhood is defined by the formula in Eq.\u0026nbsp;(\u003cspan refid=\"Equ1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$$\\:\\:{N}_{ϵ}\\left(p\\right)=\\left\\{q\\in\\:D\\left|dist\\right(p,q)\\le\\:ϵ\\right\\}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003eIn this context, D represents the dataset, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:dist(p,q)\\)\u003c/span\u003e\u003c/span\u003e denotes the distance between points \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\\)\u003c/span\u003e\u003c/span\u003e and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:q\\)\u003c/span\u003e\u003c/span\u003e。If the number of points \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{N}}_{\\text{ϵ}}\\left(p\\right)\\)\u003c/span\u003e\u003c/span\u003e within the ε-neighborhood of point p satisfies \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{N}}_{\\text{ϵ}}\\left(p\\right)\\ge\\:\\left|MinPts\\right|\\)\u003c/span\u003e\u003c/span\u003e,then point p is labeled as a core point, and a new cluster is created by including \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:p\\)\u003c/span\u003e\u003c/span\u003e and all points in \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\text{N}}_{\\text{ϵ}}\\left(p\\right)\\)\u003c/span\u003e\u003c/span\u003e; otherwise, point p is marked as noise.After completing the above operation, the algorithm selects the next unvisited point and repeats the process until all points have been visited.\u003c/p\u003e\u003cp\u003eFor DBH estimation, this study employs the least squares circle fitting method. This approach calculates the optimal circle center and radius by minimizing the sum of the squared distances from all data points to the circumference, as shown in equations (\u003cspan refid=\"Equ2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) and (\u003cspan refid=\"Equ3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Tree height is estimated using ground-referenced point clouds by performing circle fitting on the point cloud at 1.3 meters above the ground, which closely corresponds to the actual measurement height.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$$\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:\\:f\\left({x}_{a},{y}_{a},R\\right)=\\sum\\:{d}_{i}^{2}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equ3\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ3\" name=\"EquationSource\"\u003e\n$$\\:{r}_{i}=\\sqrt{{({x}_{i}-{x}_{a})}^{2}+{({y}_{i}-{y}_{a})}^{2}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e3\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:di\\)\u003c/span\u003e\u003c/span\u003e represents the distance from each point at 1.3 m height to the center of the fitted circle, defined as \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:di=({r}_{i}-R)\\)\u003c/span\u003e\u003c/span\u003e. xa and ya are the coordinates of the determined center of the fitted circle, while \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{x}_{i}\\)\u003c/span\u003e\u003c/span\u003e ​ and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{y}_{i}\\)\u003c/span\u003e\u003c/span\u003e are the coordinates of the circle center during the iterative fitting process. \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{r}_{i}\\)\u003c/span\u003e\u003c/span\u003e denotes the radius of the circle fitted to each individual point, and R is the radius of the final determined fitted circle.\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"3. Experiment","content":"\u003cp\u003eTo evaluate the effectiveness of the added and improved modules in the overall system, we designed corresponding experiments and conducted quantitative assessments from four aspects: semantic segmentation accuracy (Section 3.1), system performance (Section 3.2), 3D map reconstruction (Section 3.3), and individual tree seg-mentation accuracy (Section 3.4).\u003c/p\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e3.1. Semantic segmentation accuracy\u003c/h2\u003e\u003cp\u003eTo verify the accuracy of the semantic segmentation algorithm used in the system, we conducted evaluations on both a public dataset and a self-collected dataset. In ad-dition, several classical semantic segmentation algorithms were trained on the same datasets to serve as baselines for comparing and validating the accuracy of our pro-posed method. For classification results, we evaluate accuracy using average precision (Avg) and mean Intersection over Union (mIoU), and assess the real-time performance of the algorithms by measuring the number of point cloud scans processed per second (Scans/sec).\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section3\"\u003e\u003ch2\u003e3.1.1Experimental details\u003c/h2\u003e\u003cp\u003eThe experiments were conducted on a system running Ubuntu 18.04, equipped with an NVIDIA GXT4090 GPU and 24 GB of RAM, using Python 3.6 as the program-ming environment. After receiving the LiDAR-scanned point cloud, the system per-forms spherical projection on all points to convert the 3D point cloud into a 2D range image, which is then fed into the network for semantic segmentation. The 2D range image is first processed through the Efficient Local Attention (ELA) module, enabling the network to focus on key local regions, such as the junctions between tree trunks and foliage or between trunks and the ground. Subsequently, the image is passed through the semantic segmentation component of the SqueezeSegV3 algorithm, which consists of layers with output channel sizes of 64, 128, 256, 256, and 256, respectively. After segmentation, a 2D prediction label map is generated and then projected back into the 3D space. All points from the LiDAR scan are projected onto a 2D range image with a resolution of 64 \u0026times; 2048. PolarNet(Zhang et al. 2020) ,SalsaNet(Aksoy et al. 2020) ,Rangnet++(Milioto et al. 2019) included in the evaluation are config-ured using their default configuration files.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section3\"\u003e\u003ch2\u003e3.1.2. Accuracy evaluation experiment\u003c/h2\u003e\u003cp\u003eThe public dataset used in this study is the Plot_a dataset (Kaijaluoto et al. 2022), which covers the study area of \u0026Auml;mari, Finland (latitude 61.19\u0026deg;, longitude 25.11\u0026deg;). LiDAR data were col-lected from three test sites(A, B, and C)which were used as the training, testing, and prediction sets, respectively. The forest plot size is 32 \u0026times; 32 meters. We utilized three classes from the dataset for point cloud classification: trunk, foliage, and ground. Points belonging to other classes were excluded from the analysis.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eEvaluation of algorithm accuracy in the Plot_a dataset\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003emIoU\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eAvg\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eScans/sec\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ePolarNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.464\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.618\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e13\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSalsaNet\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.554\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.828\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e17\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRangnet++\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.481\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.851\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e12\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSqueezeSegV3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.506\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.832\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e20\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eours\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.558\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.853\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e18\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eAs shown in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, on the public dataset, our algorithm performs well on both mIoU and Avg metrics. The mIoU increases by 5% compared to the algorithm before improvement, indicating higher overall segmentation quality. The Avg is also slightly higher than other algorithms, showing that it can more accurately distinguish different categories of point clouds. In addition, while maintaining improved accuracy, our al-gorithm's point cloud processing speed is slightly lower than SqueezeSegV3, but better than other algorithms, demonstrating good real-time performance.\u003c/p\u003e\u003cp\u003eTo validate the effectiveness of the algorithm, we also collected point cloud data using a lidar-equipped unmanned vehicle in two selected 25\u0026times;25 m birch forest plots within the experimental area, which were used as the training set and the prediction set, respectively. We trained the model using the training set and validated it on the prediction set. During the annotation of the training point clouds, to facilitate subse-quent individual tree segmentation, the point clouds were classified into ground, trunk, foliage and others. Points belonging to other categories were set as ignore points dur-ing training.\u003c/p\u003e\u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eComparison of Algorithm Accuracy on the Self-Collected Dataset\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 11.2987%;\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 7.4026%;\"\u003e\n \u003cp\u003eground\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 7.0909%;\"\u003e\n \u003cp\u003efoliage\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 6.7792%;\"\u003e\n \u003cp\u003etrunk\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 6.7792%;\"\u003e\n \u003cp\u003emIoU\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\" style=\"width: 6.0779%;\"\u003e\n \u003cp\u003eAvg\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" style=\"width: 5.2208%;\"\u003e\n \u003cp\u003eScans/sec\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 14.2597%;\"\u003e\n \u003cp\u003ePolarNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.0909%;\"\u003e\n \u003cp\u003e0.684\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.7792%;\"\u003e\n \u003cp\u003e0.638\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.7792%;\"\u003e\n \u003cp\u003e0.592\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0779%;\"\u003e\n \u003cp\u003e0.537\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.013%;\"\u003e\n \u003cp\u003e0.807\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6%;\"\u003e\n \u003cp\u003e14\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 14.2597%;\"\u003e\n \u003cp\u003eSalsaNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.0909%;\"\u003e\n \u003cp\u003e0.556\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.7792%;\"\u003e\n \u003cp\u003e0.254\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.7792%;\"\u003e\n \u003cp\u003e0.781\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0779%;\"\u003e\n \u003cp\u003e0.664\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.013%;\"\u003e\n \u003cp\u003e0.848\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6%;\"\u003e\n \u003cp\u003e22\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 14.2597%;\"\u003e\n \u003cp\u003eRangnet++\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.0909%;\"\u003e\n \u003cp\u003e0.684\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.7792%;\"\u003e\n \u003cp\u003e0.210\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.7792%;\"\u003e\n \u003cp\u003e0.771\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0779%;\"\u003e\n \u003cp\u003e0.655\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.013%;\"\u003e\n \u003cp\u003e0.876\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6%;\"\u003e\n \u003cp\u003e17\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 14.2597%;\"\u003e\n \u003cp\u003eSqueezeSegV3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.0909%;\"\u003e\n \u003cp\u003e0.378\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.7792%;\"\u003e\n \u003cp\u003e0.629\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.7792%;\"\u003e\n \u003cp\u003e0.892\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0779%;\"\u003e\n \u003cp\u003e0.633\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.013%;\"\u003e\n \u003cp\u003e0.858\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6%;\"\u003e\n \u003cp\u003e27\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 14.2597%;\"\u003e\n \u003cp\u003eOurs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.0909%;\"\u003e\n \u003cp\u003e0.713\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.7792%;\"\u003e\n \u003cp\u003e0.719\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.7792%;\"\u003e\n \u003cp\u003e0.891\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 6.0779%;\"\u003e\n \u003cp\u003e0.672\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" colspan=\"2\" style=\"width: 7.013%;\"\u003e\n \u003cp\u003e0.896\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\" style=\"width: 6%;\"\u003e\n \u003cp\u003e25\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\u003cp\u003eAs shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, different semantic segmentation algorithms exhibit signifi-cant performance differences in forest environments. In terms of results, our algorithm performs slightly better overall than other models, demonstrating its strong perfor-mance in complex natural environments. Specifically, the confidence scores of our al-gorithm for the classes ground, foliage, and trunk are 0.713, 0.719, and 0.891, respec-tively. The overall mIoU reaches 0.772, and the average precision reaches 0.896, both significantly higher than those of RangeNet++, SalsaNet, PolarNet, and the original SqueezeSegV3 before improvement. This indicates that the introduced ELA attention mechanism enhances the model\u0026rsquo;s ability to perceive local spatial structures and con-textual semantics. By guiding the network to focus on key local regions in the forest point cloud, it significantly improves the segmentation accuracy of fine-grained targets. Although the point cloud processing efficiency of our algorithm is slightly lower than that of the original algorithm, it still exceeds the LiDAR acquisition rate (10 Hz), meet-ing the real-time processing requirements for point cloud data. Considering both ac-curacy and real-time performance in forest environments, the proposed algorithm is capable of handling the complexity of forest scenes effectively.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003e3.2. System Performance Evaluation\u003c/h2\u003e\u003cp\u003ePrecise localization accuracy is critical for improving the quality of map recon-struction. Therefore, we collected trajectory data at the experimental site and used RTK-recorded trajectories as ground truth. We evaluated the trajectory accuracy of both our improved algorithm and the baseline method. As shown in Table Forest_01, we used the EVO tool to compute the Root Mean Square Error (RMSE), standard devi-ation (Std), and maximum error (Max) of the trajectories.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eDue to the complexity of forest environments, terrain conditions can hinder the unmanned vehicle from performing loop closure during data collection. To address this scenario, we recorded a long-distance dataset without loop closures, named For-est_02, to evaluate the trajectory alignment accuracy of the system under loop-closure-absent conditions, as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e10\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eRMSE, Std, and Max Errors of Different Algorithms\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e\u003cp\u003eForest_01\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colspan=\"3\" nameend=\"c7\" namest=\"c5\"\u003e\u003cp\u003eTree_02\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRMSE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eStd\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMax\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eRMSE\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eStd\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003eMax\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLio_SAM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.0214\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.0118\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.554\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.0291\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.0174\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.558\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eours\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.0213\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.0110\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.773\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.0233\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.0102\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.709\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eIn the Forest_01 and Tree_02 scenarios, the trajectory accuracy comparison be-tween LIO-SAM and our proposed method reveals that the RMSE values of both algo-rithms are very close. However, our method achieves lower standard deviations (0.0110 and 0.0102), indicating reduced fluctuation in trajectory error and thus im-proved stability. Although the maximum error is slightly higher at 0.77 m, the overall system demonstrates stronger stability, offering enhanced accuracy and robustness in complex forested environments with irregular tree distributions. Overall, our approach improves error consistency while maintaining accuracy, showcasing better adaptability to challenging conditions.\u003c/p\u003e\u003cdiv id=\"Sec14\" class=\"Section3\"\u003e\u003ch2\u003e3.2.2. Evaluation of System Runtime Efficiency\u003c/h2\u003e\u003cp\u003eThis section evaluates the runtime efficiency of different algorithms when pro-cessing a single LiDAR scan. The main evaluation metrics include average processing time (Avg rate), actual memory usage (RES), and GPU utilization (GPU). The experi-mental data were collected during real-time operation in a forest environment, as shown in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e,the LIO-SAM algorithm requires a longer time to extract trunk fea-tures from raw point clouds, with an average processing time of 27.57 ms per frame and a high GPU utilization of 71.3%. In contrast, our proposed improved algorithm demonstrates significantly better performance, reducing the average runtime to 24.43 ms and lowering GPU usage to 50.5%, while maintaining a comparable memory foot-print.\u003c/p\u003e\u003cp\u003eThe improvement can be attributed to the introduction of the iKD-tree data structure during the point cloud preprocessing stage. This structure allows efficient dynamic insertion and removal of points, thereby optimizing the data processing pipe-line, reducing redundant computation, and enhancing memory utilization. Overall, our approach significantly enhances the system's real-time performance and operational efficiency.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eEfficiency of different algorithms to complete a build\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAvg rate\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eRES\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eGPU\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLio_SAM\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e27.57ms\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e268MB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e71.3%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eours\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e24.43ms\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e286MB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e50.5%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003e3.3. 3D Map Reconstruction\u003c/h2\u003e\u003cp\u003eAccurate classification of individual tree point clouds can significantly improve the precision of clustering algorithms. To verify this, we separately collected point cloud data of a row of trees and performed semantic segmentation using the trained model. We then conducted comparative experiments with other algorithms to evaluate the effectiveness of our method. The segmentation results are presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e10\u003c/span\u003e. From the segmentation results, our algorithm accurately classified the point clouds of tree trunks, canopies, and ground. It preserved the canopy boundaries well, with no significant omissions or misclassifications. In contrast, both RangeNet\u0026thinsp;+\u0026thinsp;+\u0026thinsp;and Squeez-eSegV3 exhibited issues such as canopy fragmentation, confusion between trunk and ground points, and blurred boundaries. Overall, after incorporating the ELA attention mechanism, our method effectively mitigates the misclassification of boundary points, thereby improving the segmentation accuracy and enhancing the overall robustness of the system.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo further evaluate the feasibility of the proposed system in large-scale forest en-vironments, we conducted real-time 3D semantic reconstruction in a birch forest. A test area of 25 m \u0026times; 25 m was selected, and the equipment shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e was used to perform a full loop traversal at a constant speed within the forest. To facilitate subse-quent tree counting and DBH estimation, the point cloud data were semantically seg-mented into three categories: ground, trunk, and foliage. The resulting 3D map is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e11\u003c/span\u003e.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eIt can be observed that the system effectively classifies the point cloud into three categories: foliage, trunk, and ground. After classification, the system removes foliage points, outliers, and other irrelevant points that could affect clustering accuracy, thereby retaining a clean 3D map composed mainly of trunk and ground points. This significantly improves the precision of single-tree segmentation and clustering. Alt-hough some misclassifications may occur for high-elevation canopy points, these do not impact the estimation of tree count or DBH. Overall, the reconstructed map demonstrates the system\u0026rsquo;s capability for accurate forest data acquisition and measurement.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e3.4. Evaluation of Individual Tree Segmentation Accuracy\u003c/h2\u003e\u003cp\u003eWe validated the individual tree segmentation using two datasets of different tree types collected from the experimental area. As shown on the left side of Fig.\u0026nbsp;12, the point cloud was filtered by the system\u0026rsquo;s semantic segmentation module to retain only ground and trunk points. For better visualization, each cluster is displayed in a differ-ent color.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAs shown in the clustering results of the point cloud in Fig.\u0026nbsp;12, the system ef-fectively performs trunk point cloud clustering, with each individual tree accurately segmented into a distinct cluster, and no erroneous clustering observed. Due to the relatively low segmentation accuracy of high-elevation canopy points, we introduced a Z-axis threshold during clustering: clusters whose points all exceed this height thresh-old are discarded. As illustrated in Fig.\u0026nbsp;12(b), filtering out these high points signifi-cantly improves clustering accuracy without affecting the estimation of tree count or DBH, thereby enhancing the overall precision of the system.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eTo evaluate the accuracy of DBH estimation, we manually measured 20 trees along the trajectory within the experimental area. The ground-truth DBH values were obtained by measuring the circumference of each tree at 1.3 meters above the ground and converting it to diameter. We then compared the actual DBH values with the fitted DBH values obtained from our system to assess the estimation accuracy. As shown in the Table \u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, we present a subset of the results along with the average error.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eAssessment of tree diameter at breast height measurements\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003e`\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTruth(cm)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eFitted (cm)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eDeviation(cm)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eRelative Ac\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e32.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e31.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e97.2%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e31.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e30.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e98.0%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e14.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e15.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e93.5%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e23.7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e24.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e96.3%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e23.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e23.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e97.4%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e21.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e21.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e99.1%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e16.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e17.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e97.7%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e15.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e15.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e97.5%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e17.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e16.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e95.3%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e10\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e17.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e18.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e97.2%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAvg\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e21.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e21.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.58\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e98.6%\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe collected LiDAR point cloud data were used as experimental input for the system. As shown in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, the DBH fitting results demonstrate overall good per-formance, with an average deviation of 0.58 cm. Considering the average actual DBH of 21.2 cm, the error is relatively small, resulting in an average relative accuracy of 98.6%, indicating that the model has strong fitting capability. Among the 10 samples, most fitting errors were within 1 cm, with only samples No. 1 and No. 3 exceeding 1.0 cm, yet still within the acceptable range of forestry measurement standards. In terms of relative accuracy, all but one sample achieved over 95%, with most exceeding 97%. Overall, the fitting results show high accuracy and consistency, demonstrating the system\u0026rsquo;s effectiveness for forest applications.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eIn this study, we presented a lightweight LiDAR SLAM system integrated with semantic segmentation, specifically designed to meet the demands of real-time per-formance and accurate individual tree segmentation in forest environments. To enhance real-time capability, the system incorporates the IKD-Tree algorithm, which constructs a dynamic KD-tree structure to significantly improve nearest neighbor search efficiency during feature matching. This optimization accelerates point cloud registration and feature extraction after downsampling, resulting in a 21.2% reduction in CPU usage compared to the original LIO-SAM algorithm and a 3 ms decrease in processing latency. In addition, the improved SqueezeSegV3 network integrated into the system achieved an mIoU of 0.67 in forestry scenarios, representing a 4% improvement over the original model. The accuracy of DBH measurement after clustering reached 98.6%, with an av-erage deviation of 0.58 cm, indicating that the system performs reliably in tree identi-fication, clustering, and DBH estimation, with a prediction error within 3%. Overall, compared with traditional LiDAR scanning systems used in forestry applications, our approach maintains high accuracy while significantly reducing system complexity, making it more suitable for deployment on mobile platforms in forest environments. In practical applications, the integration of semantic segmentation enables the effective removal of non-target point clouds such as weeds and shrubs, thereby improving the accuracy of trunk clustering and DBH estimation. These results demonstrate the sys-tem's robustness and practicality for forest resource assessment tasks.\u003c/p\u003e\u003cp\u003eOur research integrates a lightweight semantic segmentation network with an ef-ficient SLAM framework, providing a deployable and scalable solution for tasks such as forest resource investigation and individual tree extraction. However, the current system achieves high segmentation accuracy primarily for tree trunks beneath the canopy, while point clouds in the upper canopy layers tend to be misclassified, making accurate estimation of tree height parameters challenging. In future work, we aim to further optimize the network architecture to develop a more efficient and accurate deep learning model, and to explore effective methods for precise tree height estima-tion, thereby advancing the application of this system in autonomous forest resource surveys using unmanned vehicles.\u003c/p\u003e"},{"header":"5. Conclusion","content":"\u003cp\u003eIn this paper, we proposed a lightweight LiDAR SLAM system integrated with semantic segmentation and deployed it on an unmanned vehicle. The system incorpo-rates an improved lightweight semantic segmentation network, SqueezeSegV3, to en-hance the accuracy of individual tree segmentation, and employs a dynamic IKD-Tree structure to accelerate point cloud feature matching and data association. Experimental results demonstrate that the proposed method effectively addresses the challenges of low processing efficiency and limited segmentation accuracy in forest point cloud data. It achieves strong performance in tasks such as tree identification and DBH estimation, showing high practical value. In future work, we aim to extend the system's capabilities to tree height estimation and complex canopy structure extraction, promoting the broader application of autonomous vehicles in intelligent forest resource surveys and mobile operational platforms.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eConceptualization, S.Z. and Y.Z. (Yongxu Zhou).; methodology, S.Z.; software, Y.Z.(Yunfeng Zhao); formal analysis, B.S.; investigation, Y.Z(Yunfeng Zhao).; data curation, B.S. and J.W..; writing\u0026mdash;original draft preparation, S.Z. and Y.Z. (Yongxu Zhou); project administration, Y.Z. (Yunfeng Zhao); funding acquisition, J.W. All authors have read and agreed to the published version of the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAksoy EE, Baci S, Cavdar S (2020) SalsaNet: Fast Road and Vehicle Segmentation in LiDAR Point Clouds for Autonomous Driving. In: 2020 IEEE Intelligent Vehicles Symposium (IV). pp 926\u0026ndash;932\u003c/li\u003e\n\u003cli\u003eB\u0026aacute;rta V, Hanu\u0026scaron; J, Dobrovoln\u0026yacute; L, Homolov\u0026aacute; L (2022) Comparison of field survey and remote sensing techniques for detection of bark beetle-infested trees. Forest Ecology and Management 506:119984. https://doi.org/10.1016/j.foreco.2021.119984\u003c/li\u003e\n\u003cli\u003eBauwens S, Bartholomeus H, Calders K, Lejeune P (2016) Forest Inventory with Terrestrial LiDAR: A Comparison of Static and Hand-Held Mobile Laser Scanning. Forests 7:127. https://doi.org/10.3390/f7060127\u003c/li\u003e\n\u003cli\u003eBrede B, Calders K, Lau A, et al (2019) Non-destructive tree volume estimation through quantitative structure modelling: Comparing UAV laser scanning with terrestrial LIDAR. Remote Sensing of Environment 233:111355. https://doi.org/10.1016/j.rse.2019.111355\u003c/li\u003e\n\u003cli\u003eCai Y, Xu W, Zhang F (2021) ikd-Tree: An Incremental K-D Tree for Robotic Applications\u003c/li\u003e\n\u003cli\u003eChen SW, Nardari GV, Lee ES, et al (2020) SLOAM: Semantic Lidar Odometry and Mapping for Forest Inventory. IEEE Robot Autom Lett 5:612\u0026ndash;619. https://doi.org/10.1109/LRA.2019.2963823\u003c/li\u003e\n\u003cli\u003eEeckhaut MVD, Poesen J, Verstraeten G, et al (2007) Use of LIDAR‐derived images for mapping old landslides under forest. Earth Surf Processes Landf 32:754\u0026ndash;769. https://doi.org/10.1002/esp.1417\u003c/li\u003e\n\u003cli\u003eFadil S, Sebari I, Moulay MA, kadi KAE (2024) Modeling and spatialization of biomass and carbon stock using unmanned Aerial Vehicle Lidar (Lidar-UAV) metrics and forest inventory in cork oak forest of Maamora. Regional Science Policy \u0026amp; Practice 16:100127. https://doi.org/10.1016/j.rspp.2024.100127\u003c/li\u003e\n\u003cli\u003eFan Y, Feng Z, Shen C, et al (2020) A trunk-based SLAM backend for smartphones with online SLAM in large-scale forest inventories. ISPRS Journal of Photogrammetry and Remote Sensing 162:41\u0026ndash;49. https://doi.org/10.1016/j.isprsjprs.2020.02.006\u003c/li\u003e\n\u003cli\u003eGao Q, Kan J (2022) Automatic Forest DBH Measurement Based on Structure from Motion Photogrammetry. Remote Sensing 14:2064. https://doi.org/10.3390/rs14092064\u003c/li\u003e\n\u003cli\u003eGollob C, Ritter T, Kra\u0026szlig;nitzer R, et al (2021) Measurement of Forest Inventory Parameters with Apple iPad Pro and Integrated LiDAR Technology. Remote Sensing 13:3129. https://doi.org/10.3390/rs13163129\u003c/li\u003e\n\u003cli\u003eGollob C, Ritter T, Nothdurft A (2020) Forest Inventory with Long Range and High-Speed Personal Laser Scanning (PLS) and Simultaneous Localization and Mapping (SLAM) Technology. Remote Sensing 12:1509. https://doi.org/10.3390/rs12091509\u003c/li\u003e\n\u003cli\u003eHao Y, Widagdo FRA, Liu X, et al (2020) Individual Tree Diameter Estimation in Small-Scale Forest Inventory Using UAV Laser Scanning. Remote Sensing 13:24. https://doi.org/10.3390/rs13010024\u003c/li\u003e\n\u003cli\u003eHe GB, Li LL (2020) RESEARCH AND APPLICATION OF LIDAR TECHNOLOGY IN CADASTRAL SURVEYING AND MAPPING. Int Arch Photogramm Remote Sens Spatial Inf Sci XLIII-B1-2020:33\u0026ndash;37. https://doi.org/10.5194/isprs-archives-XLIII-B1-2020-33-2020\u003c/li\u003e\n\u003cli\u003eKaijaluoto R, Kukko A, El Issaoui A, et al (2022) Semantic segmentation of point cloud data using raw laser scanner measurements and deep neural networks. ISPRS Open Journal of Photogrammetry and Remote Sensing 3:100011. https://doi.org/10.1016/j.ophoto.2021.100011\u003c/li\u003e\n\u003cli\u003eKukko A, Kaijaluoto R, Kaartinen H, et al (2017) Graph SLAM correction for single scanner MLS forest data under boreal forest canopy. ISPRS Journal of Photogrammetry and Remote Sensing 132:199\u0026ndash;209. https://doi.org/10.1016/j.isprsjprs.2017.09.006\u003c/li\u003e\n\u003cli\u003eLi J, Liu J, Huang Q (2023) PointDMM: A Deep-Learning-Based Semantic Segmentation Method for Point Clouds in Complex Forest Environments. Forests 14:2276. https://doi.org/10.3390/f14122276\u003c/li\u003e\n\u003cli\u003eLiu B, Huang H, Su Y, et al (2022) Tree Species Classification Using Ground-Based LiDAR Data by Various Point Cloud Deep Learning Methods. Remote Sensing 14:5733. https://doi.org/10.3390/rs14225733\u003c/li\u003e\n\u003cli\u003eMa Z, Dong Y, Zi J, et al (2023) Forest-PointNet: A Deep Learning Model for Vertical Structure Segmentation in Complex Forest Scenes. Remote Sensing 15:4793. https://doi.org/10.3390/rs15194793\u003c/li\u003e\n\u003cli\u003eMilioto A, Vizzo I, Behley J, Stachniss C (2019) RangeNet ++: Fast and Accurate LiDAR Semantic Segmentation. In: 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp 4213\u0026ndash;4220\u003c/li\u003e\n\u003cli\u003eMu X, Zhao D, Zheng Z, et al (2025) Can Stereoscopic Density Replace Planar Density for Forest Aboveground Biomass Estimation? A Case Study Using Airborne LiDAR and Landsat Data in Daxing\u0026rsquo;anling, China. Remote Sensing 17:1163. https://doi.org/10.3390/rs17071163\u003c/li\u003e\n\u003cli\u003eNeuville R, Bates JS, Jonard F (2021) Estimating Forest Structure from UAV-Mounted LiDAR Point Cloud Using Machine Learning. Remote Sensing 13:352. https://doi.org/10.3390/rs13030352\u003c/li\u003e\n\u003cli\u003eNuchter A, Lingemann K, Hertzberg J (2007) Cached k-d tree search for ICP algorithms. In: Sixth International Conference on 3-D Digital Imaging and Modeling (3DIM 2007). pp 419\u0026ndash;426\u003c/li\u003e\n\u003cli\u003eOveland I, Hauglin M, Giannetti F, et al (2018) Comparing Three Different Ground Based Laser Scanning Methods for Tree Stem Detection. Remote Sensing 10:538. https://doi.org/10.3390/rs10040538\u003c/li\u003e\n\u003cli\u003ePierzchała M, Gigu\u0026egrave;re P, Astrup R (2018) Mapping forests using an unmanned ground vehicle with 3D LiDAR and graph-SLAM. Computers and Electronics in Agriculture 145:217\u0026ndash;225. https://doi.org/10.1016/j.compag.2017.12.034\u003c/li\u003e\n\u003cli\u003eQian C, Liu H, Tang J, et al (2016) An Integrated GNSS/INS/LiDAR-SLAM Positioning Method for Highly Accurate Forest Stem Mapping. Remote Sensing 9:3. https://doi.org/10.3390/rs9010003\u003c/li\u003e\n\u003cli\u003eRajab Pourrahmati M, Baghdadi N, Fayad I (2023) Comparison of GEDI LiDAR Data Capability for Forest Canopy Height Estimation over Broadleaf and Needleleaf Forests. Remote Sensing 15:1522. https://doi.org/10.3390/rs15061522\u003c/li\u003e\n\u003cli\u003eShan T, Englot B, Meyers D, et al (2020) LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping\u003c/li\u003e\n\u003cli\u003eTang J, Chen Y, Kukko A, et al (2015) SLAM-Aided Stem Mapping for Forest Inventory with Small-Footprint Mobile LiDAR\u003c/li\u003e\n\u003cli\u003eTian L, Wu X, Tao Y, et al (2023) Review of Remote Sensing-Based Methods for Forest Aboveground Biomass Estimation: Progress, Challenges, and Prospects. Forests 14:1086. https://doi.org/10.3390/f14061086\u003c/li\u003e\n\u003cli\u003eWang Y, Lin Y, Cai H, Li S (2022) Hierarchical Fine Extraction Method of Street Tree Information from Mobile LiDAR Point Cloud Data. Applied Sciences 13:276. https://doi.org/10.3390/app13010276\u003c/li\u003e\n\u003cli\u003eXu C, Wu B, Wang Z, et al (2021a) SqueezeSegV3: Spatially-Adaptive Convolution for Efficient Point-Cloud Segmentation\u003c/li\u003e\n\u003cli\u003eXu D, Wang H, Xu W, et al (2021b) LiDAR Applications to Estimate Forest Biomass at Individual Tree Scale: Opportunities, Challenges and Future Perspectives. Forests 12:550. https://doi.org/10.3390/f12050550\u003c/li\u003e\n\u003cli\u003eXu W, Wan Y (2024) ELA: Efficient Local Attention for Deep Convolutional Neural Networks\u003c/li\u003e\n\u003cli\u003eYang S, Xing Y, Xing T, et al (2024) Multisensors Fusion SLAM-Aided Forest Plot Mapping With Backpack Dual-LiDAR System. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 17:16051\u0026ndash;16070. https://doi.org/10.1109/JSTARS.2024.3451175\u003c/li\u003e\n\u003cli\u003eYu J, Lei L, Li Z (2024) Individual Tree Segmentation Based on Seed Points Detected by an Adaptive Crown Shaped Algorithm Using UAV-LiDAR Data. Remote Sensing 16:825. https://doi.org/10.3390/rs16050825\u003c/li\u003e\n\u003cli\u003eZhang T, Long J, Lin H, et al (2024) A Novel Feature Evaluation Method in Mapping Forest AGB by Fusing Multiple Evaluation Metrics Using PolSAR Data. IEEE Geoscience and Remote Sensing Letters 21:1\u0026ndash;5. https://doi.org/10.1109/LGRS.2024.3378425\u003c/li\u003e\n\u003cli\u003eZhang Y, Zhou Z, David P, et al (2020) PolarNet: An Improved Grid Representation for Online LiDAR Point Clouds Semantic Segmentation. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, Seattle, WA, USA, pp 9598\u0026ndash;9607\u003c/li\u003e\n\u003cli\u003eZhao X, Guo Q, Su Y, Xue B (2016) Improved progressive TIN densification filtering algorithm for airborne LiDAR data in forested areas. ISPRS Journal of Photogrammetry and Remote Sensing 117:79\u0026ndash;91. https://doi.org/10.1016/j.isprsjprs.2016.03.016\u003c/li\u003e\n\u003cli\u003eZhou T, Zhao C, Wingren CP, et al (2024) Forest feature LiDAR SLAM (F2-LSLAM) for backpack systems. ISPRS Journal of Photogrammetry and Remote Sensing 212:96\u0026ndash;121. https://doi.org/10.1016/j.isprsjprs.2024.04.025\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"annals-of-forest-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Annals of Forest Science](https://link.springer.com/journal/13595)","snPcode":"13595","submissionUrl":"https://submission.springernature.com/new-submission/13595/3","title":"Annals of Forest Science","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"3D LiDAR, Forest, Point cloud, Semantic segmentation","lastPublishedDoi":"10.21203/rs.3.rs-6971887/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6971887/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cem\u003e\u003cstrong\u003eContext\u003c/strong\u003e\u003c/em\u003e:\u003c/p\u003e\n\u003cp\u003eEfficient and automated acquisition of individual tree parameters is essential for intelligent forest resource inventory. Traditional methods, which rely heavily on manual measurements, are limited in their scalability and cannot meet the demands of large-scale, high-frequency surveys.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eAims\u003c/strong\u003e\u003c/em\u003e:\u003c/p\u003e\n\u003cp\u003eThis study aims to develop a lightweight LiDAR SLAM system integrated with semantic segmentation to improve the accuracy and real-time performance of individual tree extraction and DBH measurement in forest environments.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eMethods:\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe system is based on an improved LIO-SAM framework, incorporating the IKD-Tree to enhance efficiency. In the semantic segmentation module, the SqueezeSegV3 network with the added ELA attention mechanism is employed to improve semantic category recognition. The segmented point clouds are processed through clustering and cylindrical fitting to extract individual trees and estimate their DBH.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/em\u003e:\u003c/p\u003e\n\u003cp\u003eOn both a public dataset and field-collected forest data, the semantic segmentation achieved accuracies of 0.85 and 0.89, with mean Intersection over Union values of 0.55 and 0.67, respectively. The average prediction accuracy for DBH reached 98.6%..\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cstrong\u003eConclusion\u003c/strong\u003e\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe system integrates a lightweight semantic network with an efficient point cloud structure, offering both high accuracy and real-time performance. It meets the requirements of large-scale and high-efficiency measurements in forest resource inventory tasks.\u003c/p\u003e","manuscriptTitle":"A Lightweight LiDAR SLAM System Integrated with Semantic Segmentation for Forest Biomass Parameter Acquisition","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-07-10 10:30:39","doi":"10.21203/rs.3.rs-6971887/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-09-11T09:34:37+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-09T01:38:06+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-03T00:51:38+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"88513190480177047495952756155962170808","date":"2025-08-27T13:12:12+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"195727378393708200879151333086620939471","date":"2025-08-27T01:48:19+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-15T08:50:09+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"201840971681869236925612040994688466875","date":"2025-08-09T22:20:44+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-07-08T13:16:21+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-07-03T12:38:43+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-07-01T16:51:33+00:00","index":"","fulltext":""},{"type":"submitted","content":"Annals of Forest Science","date":"2025-06-25T07:35:11+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"annals-of-forest-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [Annals of Forest Science](https://link.springer.com/journal/13595)","snPcode":"13595","submissionUrl":"https://submission.springernature.com/new-submission/13595/3","title":"Annals of Forest Science","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"e4b32524-0e0e-4e5b-89db-152bc705a3cc","owner":[],"postedDate":"July 10th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-07T09:55:55+00:00","versionOfRecord":[],"versionCreatedAt":"2025-07-10 10:30:39","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6971887","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6971887","identity":"rs-6971887","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}