Tree Species Identification Of UAV Multi-source Data Based On Deep Learning

Accurate identification of forest tree species has significant significance for the refined management of forestry.Unmanned aerial vehicles(UAV)visible light images contain rich texture and boundary information,while point cloud data contain rich spatial information.The joint use of these two UAV remote sensing data,playing to their respective advantages,is beneficial to improving the tree species recognition rate.Research on tree species recognition based on deep learning has shown that its recognition results are generally superior to traditional algorithms.However,existing deep learning models still have some problems in tree species classification based on UAV data.For example,the ability to extract shallow features such as the canopy boundary in images is insufficient,and the ability to extract global features of sparse canopy point clouds is limited.In addition,the universality of the method is low,data segmentation is difficult,and dataset annotation is time-consuming and labor-intensive.To solve these problems,this paper proposes a deep learning tree species recognition solution suitable for UAV multisource remote sensing data,with the following main research content and innovative points:(1)Multidimensional crown dataset construction.In response to the time-consuming issue in constructing image datasets,the use of superpixel segmentation instead of uniform cropping was proposed.The segmented pixel blocks mostly have intact tree crown boundaries,reducing the fragmentation caused by uniform cropping and aiding in data annotation.Experimental results show that this method not only saves 35% of annotation time but also improves the segmentation accuracy of Mask-RCNN by 2.19%.To address the difficulty of single-tree crown segmentation in point cloud datasets,a segmentation method based on two-dimensional tree crown masks was proposed.By using an instance segmentation network to extract the image tree crown segmentation mask and mapping it to the point cloud XY plane,automatic segmentation and annotation of corresponding tree crown areas was achieved,and the resulting point cloud of segmented crowns fused the boundary information of the two-dimensional tree crowns,achieving dimensionality reduction in the point cloud single-tree segmentation.Experimental results show that the dataset constructed using this approach has the highest match rate and overall accuracy with manually segmented datasets,reaching 97.69% and 91.81% respectively,and improves the tree species identification rate of Point Net++ by 1.4%,providing a reference for the extraction of specific features from point cloud using images.(2)Instance segmentation of tree crowns with UAV images.A new network structure is proposed to address the lack of attention to contextual information at the boundaries of tree crowns and insufficient extraction of shallow features in Mask-RCNN.The structure incorporates Atrous Spatial Pyramid Pooling(ASPP)to expand the receptive field and improve the network’s attention to contextual information.It also includes a Convolutional Block Attention Module(CBAM)to enhance the network’s attention to shallow features and mitigate the feature discontinuity caused by atrous convolution.Experimental results show that the improved network achieved an overall accuracy of 93.19% for tree crown segmentation,with an increase of 4.87%.The segmentation performance of sparse forests was better than that of dense forests,with accuracies of 94.54% and 91.48%,respectively,which were better than other traditional algorithms.The network exhibited high robustness for satellite remote sensing images and other UAV images,and can be specially used for tree crown segmentation tasks with UAV data.(3)Tree species identification using drone point cloud data.In order to address the issue of insufficient global feature extraction ability of sparse tree crown point clouds using Point Net++,a new network structure is proposed.The network is assisted in training by introducing tree crown convex hulls as additional input features,which enhances the model’s ability to learn tree crown shape features.The network also employs a multi-neighborhood feature mechanism that aggregates features between points within and between the neighborhoods of the center point,achieving mixed multi-neighborhood features,thus expanding the receptive field and enhancing the network’s ability to extract global features from sparse point clouds.Experimental results show that the overall recognition rate of the improved network for tree species is 87.7%,an increase of 5%.Furthermore,when the convex hull features and multi-neighborhood features are separately introduced,the recognition rate is increased by 2.9% and 1.8%,respectively.Convex hull features significantly improve identification accuracy for data with regular appearances,up to 3.4%,while multi-neighborhood features are sensitive to sparse and scattered point cloud data and improve recognition rate up to 1.6%.Additionally,the model achieved higher identification rates for broad-leaved trees than for coniferous trees,with recognition rates of 90.8% and 87.9%,respectively.