Segmentation of individual mangrove trees using UAV-based LiDAR data

Abstract

Accurate assessment of structural parameters is essential to effectively monitor the mangrove resources. However, the extraction results of mangrove structural parameters are closely related to the segmentation results of individual trees. Although the results of individual tree segmentation are influenced by many factors, the specific factors affecting the segmentation results of individual mangrove trees, such as data source, image resolution, segmentation algorithm, and stand density, have not yet been elucidated. Therefore, in this study, canopy height models (CHMs) with different spatial resolutions were derived from unmanned aerial vehicle (UAV)-based light detection and ranging (LiDAR) data. Moreover, the watershed algorithm (WA), regional growth (RG), and improved K-nearest neighbour (KNN) and bird's eye view (BEV) faster region-based convolutional neural network (R-CNN) algorithms were used to segment the individual mangrove trees based on CHMs and LiDAR data at three sites with varying stand densities. Finally, different segmentation algorithms, image resolutions, and forest densities were comparatively assessed to determine their influence on the segmentation results of individual trees. Segmentation accuracy of the improved KNN algorithm was the highest among the CHM-based algorithms, such as the WA, RG, and improved KNN algorithms, with an optimal F of 0.893 and minimum F of 0.628. R-CNN algorithm based on LiDAR data had an optimal F value of 0.931 and minimum F value of 0.612. Based on the segmentation results, the overall accuracy ranking of the different segmentation algorithms was BEV Faster R-CNN > improved KNN > RG > WA. The ranking of the segmentation results for sites with different stand densities was low-density (LD) > medium-density (MD) > high-density (HD). For LD and MD sites, the BEV Faster R-CNN algorithm had the highest F values (0.931 and 0.712, respectively). For the HD site, all algorithms performed poorly, and the F values of all algorithms, except the RG algorithm, were higher than 0.6. Based on the segmentation results of different spatial resolutions, CHM result with 0.1 m was the best, being better than the CHM results with 0.25 and 0.5 m. Our results demonstrated that all segmentation algorithms, spatial resolutions, and stand densities affected the segmentation results for individual mangrove trees. Although the segmentation results of the deep learning algorithm were better than those of the other algorithms, the segmentation results at the HD site were limited. Therefore, further research is necessary to improve the accuracy of the segmentation results for individual mangrove trees at HD sites.