Urban street trees play a significant ecological role in urban environment enhancement, air purification, noise reduction and shading. The survey of street trees constitutes a vital part of the survey of urban ecological resources. The manual survey method of street trees in traditional cities is disadvantageous because of its high cost and low efficiency, making it hard to satisfy the needs of administrative departments to monitor urban ecological information efficiently. Although extracting the information on urban street trees by means of low-altitude remote sensing of UAV can boost the efficiency of field surveys, there are still some problems such as excessive workload in the manual editing of DEM and segmentation of tree crowns during the processing of industry data. To overcome the above shortcomings, this paper probed into a scheme for the information extraction of urban street trees by remote sensing of UAV based on the instance segmentation model of improved Mask R-CNN.

Based on the deep learning of the segmentation Mask R-CNN model, this paper used the non-maximum suppression NMS (Non-Maximum Suppression) algorithm to suppress all detected objects in the absence of complete overlap of urban street trees and determined whether the detection was repeated according to the overlap ratio between the detection canopy area and the mask to be detected, so as to improve the data training efficiency and accuracy. The accuracy evaluation index was constructed to evaluate the performance and adaptability of the model according to the prediction results of the test set.

Using the same DJI Phantom 4pro drone equipped, image data collection was carried out on the typical urban street trees such as camphor trees, French plane trees, silver birch, etc. to create a street tree dataset based on UAV remote sensing images, and the Mask R-CNN segmentation model was used to detect and segment the canopy of street trees, and compare and analyzes with manual extraction methods. The results showed that the average overall accuracy of automatic canopy segmentation of street trees by this method was 0.865, the average detection rate was 0.887, and after simple manual correction, the average overall accuracy of canopy segmentation of street trees can reach 0.948 and the average detection rate can reach 0.965.

Aerial survey of the field data by UAV is advantageous in short acquisition cycles and high efficiency. It can automatically acquire relatively accurate information on crown detection and contour segmentation when combined with the improved Mask R-CNN instance segmentation algorithm for deep learning, which saves a great deal of time for field survey and industry data processing, showing enormous significance for boosting the monitoring efficiency and the automation level of urban street trees.

Unmanned aerial vehicle (UAV) mounted LiDAR technology offers advantages such as high flexibility, short monitoring cycles, and the ability to penetrate through obstacles. In practical applications, due to the occlusion between tree trunks, it is easy to result in insufficient point cloud data density for the canopy obtained by LiDAR. Moreover, the algorithms for extracting tree trunks from LiDAR point clouds are still immature, with challenges such as large data processing volumes and low extraction accuracy. This study aims to improve the accuracy and efficiency of extracting tree trunks from UAV LiDAR data to provide accurate and efficient technical support for forest monitoring and forestry management.

The UAV carried LiDAR to fly multiple times at different altitudes to obtain complete point clouds covering the canopy and the understory. A classification method based on echo types was used to divide the point cloud data into ground points and vegetation points, and a digital terrain model was generated using the ground point cloud. Based on the difference in reflection intensity between tree trunks and leaves, classification was performed using thresholds and semi-supervised support vector machine algorithms, and point clusters were segmented into connected areas using the region growing method. The intersections of the connected areas with the horizontal plane were calculated, and the intersection information at the same height was integrated to reconstruct the tree trunk model. Hypothesis testing was used to judge the continuity of the cross-section and the variation in radius, to identify and construct a complete tree trunk mesh model.

A multi-rotor UAV equipped with LiDAR was used to collect canopy point cloud data at an altitude of 50 m, with a 90% flight direction overlap rate, a 75% lateral overlap rate, and a gimbal lens angle of 90°, at a flight speed of 4 m/s. At the same time, the same equipment was used to obtain more refined point cloud information under the canopy at heights of 1.5-2.5 m, with gimbal angles of 0°, 45°, and -45°. After processing, the vegetation and ground point clouds were removed, and the tree trunk point cloud was accurately extracted. The use of region growing to segment the point cloud and generate a two-dimensional distance image, followed by noise filtering and color segmentation, further improved the data quality. The application of the line set and RANSAC algorithm effectively estimated the shape of the main trunk cross-section and constructed an accurate tree trunk model. Testing on a Lenovo ThinkStation graphics workstation showed that the extraction of the tree trunk point cluster took only 15.9 seconds, and the generation of the tree trunk mesh model also took only 71.5 seconds. The average accuracy of the extracted breast diameter of the tree trunk was 0.958, and the average extraction accuracy of the tree height reached 0.964.

The article applied a multi-rotor UAV equipped with LiDAR, using a multi-flight, multi-altitude, and multi-attitude method to collect high-quality forest point cloud data. By combining echo-type point cloud data classification, region growing, distance image construction, noise filtering, and color segmentation, the single tree trunk point cloud segmentation algorithm was improved, effectively estimating the shape of the main trunk cross-section, and constructing an accurate tree trunk model. This approach overcomes the limitations of canopy penetration rate and trunk occlusion, effectively improving the efficiency and accuracy of tree trunk extraction. It provides high-precision tree trunk data for forest resource management and has significant practical application value for improving the efficiency and automation level of forest monitoring.