【Objective】Rapid and accurate identification of tree species is crucial for the research and protection of forest resources. Identification of tree species by remote sensing technology has become an important method of forest investigation. However, there are some problems in tree species identification by remote sensing, such as a lack of information in the upper canopy of terrestrial laser scanning (TLS) data and the lower canopy of unmanned aerial vehicle (UAV) images. Therefore, identifying tree species requires multi-source remote sensing data.【Method】In this study, Pinus sylvestris var. mongolica, P. tabuliformis var. mukdensis, Fraxinus mandshurica and Juglans mandshurica in the urban forestry demonstration base of Northeast Forestry University at Harbin were used as the research objects to identify tree species. TLS point cloud data and UAV image data were acquired. Through the processing of UAV images, the photogrammetric point cloud and orthophoto image were obtained. The UAV image point cloud and TLS point cloud data were registered and fused, then divided into single tree points. Based on the single tree point cloud, shape features, structure features, tree trunk color features, and crown color features were extracted, and tree species identification was performed by a support vector machine classification algorithm. Subsequently, the ability of the method to identify tree species using different characteristics was analyzed by random forest algorithm.【Result】Optimal results were obtained when all the features were used to identify tree species. The total accuracy and Kappa coefficient of tree species identification results were 93.48% and 0.91, respectively, which were improved by 4.35-16.31percentage points and 0.06-0.22, respectively, compared with other comparison schemes.【Conclusion】The tree species recognition method based on the fusion of TLS data and UAV image point cloud data proposed in this study can compensate for the lack of information in the upper canopy of TLS point cloud data and the lower canopy of UAV images to a certain extent, and make full use of the rich information contained in multi-source data for tree species recognition. The method can effectively improve the accuracy of tree species identification.
【Objective】To solve the problem that a canopy height model (CHM) and normalized point cloud (NPC) directly generated by terrestrial laser scanning (TLS) are not capable of detecting individual trees in complex stands, this study introduced the method of point cloud slicing combined with clustering to improve the detection accuracy.【Method】In this study, six sample plots in a plantation with different stand densities in Guangxi Zhuang Autonomous Region, China, were used as the research objects. First, the NPC data of a sample plot obtained by TLS were used to extract point cloud slices at a height of 1.3 m, and then the density-based spatial clustering of applications with noise (DBSCAN) and mean shift algorithms were used to cluster the tree trunk point clouds in the slices. The accuracy was verified by the field survey data, and the detection results were compared with those of the local maximum algorithm based on a CHM, and a point cloud segmentation algorithm based on an NPC. The applicability and parameter sensitivity of the different detection methods were evaluated and analyzed.【Result】Satisfactory detection results were obtained by all methods, and the optimal detection accuracy F-score was ≥ 0.86 for each sample plot. The individual tree detection method using point cloud slicing combined with a clustering algorithm produced better results. The clustering threshold epsilon neighborhood (Eps) value of the DBSCAN algorithm and the clustering radius r of the mean shift algorithm significantly affected the individual tree detection rate, with the maximum Eps depending on the maximum stand spacing and optimum results when r was close to the maximum individual tree diameter at breast height.【Conclusion】Individual tree detection based on point cloud slicing combined with a clustering algorithm can increase the detection rate of understory trees lower forest, effectively improve the accuracy of single tree detection in dense stands, and provide a reference for the selection of single tree detection methods in different forest stands.
【Objective】Due to the complex structure and features, traditional classification models for tree branches and leaf point clouds typically face several problems, including poor stability, low accuracy, model overfitting, and high computational costs. In this study, we propose an improved CART (classification and regression tree) model for leaf and branch classification based on Quercus glauca individual tree point cloud data from terrestrial laser LiDAR.【Method】First, the feature descriptor was constructed according to the neighborhood points, and the optimal value of the neighborhood search parameter was then determined. The CART model was improved by gradually introducing variables and adjusting the structure of the decision tree. The classification results of the improved CART model were compared with those of the Logistics regression and K-nearest neighbor (KNN) models.【Result】The accuracy of the improved CART model using the test data increased after introducing the feature descriptors as variables, exceeding that of the Logistics regression and KNN model by 13.1% and 13.6%, respectively. Moreover, the improved CART model exhibited higher accuracy, better stability, and marked reduced model size following the improvement. In particular, the model size was reduced by 99.9% compared with before the improvement, while the data training time was only 51.3% of that before the adjustment. The comprehensive evaluation index of the improved CART model was approximately 0.9 on both trunk and leaf data, with the difference between accuracy on train data and test data lower than 0.001, indicating no overfitting.【Conclusion】The improved CART model has a high accuracy and stability, and achieves good classification results on small samples. This study provides a methodological reference for the accurate and rapid classification of trunk and leaf point clouds from terrestrial laser LiDAR.
【Objective】Using open source active and passive remote sensing data (Sentinel-1 & -2) combined with forest canopy height can improve the estimation accuracy of forest net primary productivity (NPP), to provide a scientific basis for the formulation of forest precision management stragtegies and a “carbon peaking and carbon neutrality” strategy.【Method】Zixing City, which is a key forest area in the south, was used as the research area. Based on Sentinel-1 and Sentinel-2 active and passive remote sensing data, four models, namely multiple stepwise regression, artificial neural network, K-nearest neighbor, and random forest, were used to estimate NPP. On this basis, the canopy height obtained by Sentinel-1 through the difference between InSAR and SRTM DEM was added to analyze its effect on the accuracy of NPP estimation.【Result】(1) The mean value of forest NPP in the study area in 2019 was 7.79 t/hm2, showing the spatial distribution characteristics of high in the central southwest and low in the northwest. (2) Among the four models, the accuracy of NPP estimation by active and passive remote sensing was higher than that by single remote sensing; the accuracy of random forest estimation of regional forest NPP was the highest, and the model performed the best. (3) Adding canopy height improved the estimation accuracy of forest NPP to a certain extent, with R2 increased from 0.70 to 0.75.【Conclusion】The accuracy of NPP estimation can be improved based on Sentinel active and passive remote sensing data and canopy height factor, which is obtained by DEM difference method.
【Objective】Aiming at meeting the needs of real-time online segmentation of target in the target application technology of street trees, an online real-time segmentation method of street tree target point cloud based on mobile laser scanning (MLS) was studied, and a point cloud instance segmentation algorithm that can accurately segment the target point cloud of street trees in real time and online was established.【Method】A street tree on one side of a 300 m long street was used as the research object. By establishing a FIFO (first input, first output) buffer, several frames of three-dimensional street point cloud data collected by MLS were read at regular intervals. The street point cloud data in the FIFO buffer after reading was converted into a three-channel street image, and the street image was segmented by image instance segmentation model to obtain street tree candidate instances. Subsequently, the street tree candidate instances were fused with the detected street tree instances, the integrity of the detected street tree instances was determined, and image-point cloud mapping was performed on the detected complete street tree instances to obtain the street tree point cloud instances. Finally, threshold filtering and K-nearest neighbor (KNN) were used to optimize the point cloud instances of street trees facing point clouds.【Result】When the threshold filter parameter was set to 0.65 m and the radius parameter of KNN was set to 0.5 m, the segmentation results of the street tree target point cloud instance were optimal, with an accuracy rate of 0.986 5, a recall rate of 0.940 7, an F1 score of 0.957 6, and an average segmentation time of each frame of 5.261 ms.【Conclusion】The online segmentation method of street tree target cloud proposed in this study is effective and meets the requirements of real-time online segmentation of street tree targets.
【Objective】This study proposes the feature segmentation method Tree-DeepLab for unmanned aerial vehicle (UAV) forest images, based on an attention mechanism and the DeepLabV3+ semantic segmentation network, to extract the main feature distribution information in forest areas.【Method】First, the forest images were annotated according to feature types from six categories (Platanus orientalis, Ginkgo biloba, Populus sp., grassland, road, and bare ground) to obtain the semantic segmentation datasets. Second, the following improvements were made to the semantic segmentation network: (1) the Xception network, the backbone of the DeepLabV3+ semantic segmentation network, was replaced by ResNeSt101 with a split attention mechanism; (2) the atrous convolutions of different dilation rates in the atrous spatial pyramid pooling were connected using a combination of serial and parallel forms, while the combination of the atrous convolution dilation rates was simultaneously changed; (3) a shallow feature fusion branch was added to the decoder; (4) spatial attention modules were added to the decoder; and (5) efficient channel attention modules were added to the decoder.【Result】Training and testing were performed based on an in-house dataset. The experimental results revealed that the Tree-DeepLab semantic segmentation model had mean pixel accuracy (mPA) and mean intersection over union (mIoU) values of 97.04% and 85.01%, respectively, exceeding those of the original DeepLabV3+ by 4.03 and 14.07 percentage points, respectively, and outperforming U-Net and PSPNet.【Conclusion】The study demonstrates that the Tree-DeepLab semantic segmentation model can effectively segment UAV aerial photography images of forest areas to obtain the distribution information of the main feature types in forest areas.
