Research On Individual Tree Structure Parameter Estimation Of Mangrove Based On UAV LiDAR

Mangrove is a rare woody viviparous plant growing in the intertidal zone of tropical and subtropical coast.It has both terrestrial and marine ecological characteristics,is one of the most important coastal ecosystems in the world.The woody biota of tidal flat wetland composed of mangrove plants provides a natural habitat for a large number of fish,crabs,birds and other organisms looking for food,shelter,breeding and care areas,and protects the coastline from floods,the invasion of waves and storms,plays an important role in the environmental protection,ecological balance and biodiversity protection of the coastal zone.Mangrove structural parameters,such as tree height and crown width,are the basic quantitative parameters of mangrove growth.They are very important for describing mangrove growth,phenological information and death.They are also the basic data for monitoring and protecting the dynamic changes of mangrove resources.Therefore,the high-precision estimation of mangrove structural parameters is of great significance for the efficient development and implementation of mangrove resource dynamic change monitoring and protection.In this paper,Gaoqiao Mangrove Reserve in Lianjiang City,Guangdong Province was took as the research area,and the UAV lidar discrete point cloud data was took as the basic data.Firstly,the original point cloud data were processed by denoising,overlapping point removal,classification and point cloud normalization.The canopy height model(CHM)with different resolution(0.5m,0.25 m,0.1M)was generated by interpolation algorithm,and the relevant parameters such as point cloud height,intensity and density were extracted.Four individual tree segmentation algorithms(watershed algorithm,regional growth algorithm,k-nearest neighbor algorithm(KNN)and regions with CNN features(R-CNN)algorithm were used for mangrove individual tree segmentation respectively,and the influence of CHM resolution on individual tree segmentation accuracy is studied.Then,the random forest,support vector machine and decision tree algorithm were used to identify mangrove individual tree species respectively,so as to select the better mangrove identification algorithm.Finally,based on individual tree segmentation,the mangrove tree height and crown width estimation model is established by using the field measurement data,so as to realized the estimation of mangrove individual tree structural parameters.The research results and main conclusions are as follows:(1)The individual tree segmentation results are closely related to the stand density.On the whole,the segmentation results of low stand density are better than that of medium stand density and better than that of high stand density.In the case of low stand density and medium stand density,the individual tree segmentation algorithm based on R-CNN has the highest accuracy,F is 0.931 and 0.712 respectively,followed by the regional growth algorithm,F is0.891 and 0.673 respectively,followed by KNN,F is 0.893 and 0.633 respectively.Watershed algorithm has the lowest accuracy of individual tree segmentation,F is 0.852 and 0.660 respectively.In the case of high stand density,the accuracy of mangrove individual tree segmentation is on the contrary.The segmentation result of watershed algorithm is better than KNN,better than regional growth algorithm,and better than R-CNN.The results show that R-CNN algorithm is more suitable for mangrove individual tree segmentation with low stand density and medium stand density,while watershed algorithm is more suitable for mangrove individual tree segmentation with high stand density.(2)CHM spatial resolution has different effects on mangrove individual tree segmentation results.Among them,the segmentation accuracy of 0.25 m CHM individual tree is the highest,F is 0.830,while the segmentation accuracy of 0.5m CHM individual tree is the lowest,and the lowest F is 0.156.The results show that CHM spatial resolution has an impact on mangrove individual tree segmentation,but not the higher the resolution,the higher the segmentation accuracy of individual tree.Therefore,in the future research,we should select the appropriate spatial resolution for individual tree segmentation to obtain the highest individual tree segmentation accuracy.(3)Random forest,support vector machine and decision tree algorithm based on LiDAR point cloud height and intensity parameters can realize the recognition of mangrove individual tree species,and the overall classification accuracy is more than 84%.Among them,the classification accuracy of mangrove individual tree species based on random forest algorithm is the highest,with an overall accuracy of 92.43%,kappa coefficient of 0.91,followed by support vector machine,with an overall accuracy of 87.71%,Kappa coefficient of 0.85,and the classification accuracy of decision tree is the worst,the overall accuracy is 84.14%,and the Kappa coefficient is 0.80.In the classification results of mangrove tree species,there are differences in the classification accuracy of different tree species.Among them,the classification accuracy of Apetalous sea mulberry tree species is the highest.In the three classification algorithms,the user accuracy is more than 94%,and the product accuracy is more than 97%,while the classification accuracy of Tung flower tree species is the worst.In the three classification algorithms,the user accuracy is less than 80%,and the worst is59.09%,indicating that the misclassification of Tung flower tree species is serious.The results show that the identification of mangrove individual tree species can be realized based on UAV LiDAR point cloud.In the future research,we should focus on the parameters and algorithms suitable for the classification of Tung flower tree species,so as to further improve the accuracy of mangrove tree species classification.(4)Different individual tree segmentation algorithms have the same trend of tree overestimation accuracy under different stand densities,while the trend of crown estimation accuracy of different individual tree segmentation algorithms with different stand densities is inconsistent,but the crown area estimation accuracy also improves with the improvement of CHM spatial resolution.For tree overestimation,the estimation accuracy of the four individual tree segmentation methods is consistent: low stand density > high stand density >medium stand density;for the estimation of crown width,in the estimation results of watershed algorithm,low stand density > high stand density > medium stand density;in the estimation results of regional growth algorithm,medium stand density > high stand density >low stand density;in KNN algorithm estimation results,low stand density > medium stand density > high stand density;in the estimation results of R-CNN algorithm,low stand density > medium stand density > high stand density.