Classification Of Single Tree Species Combining Airborne CCD Images And LiDAR Point Cloud Data

Tree species identification is of great significance to the management and sustainable development of forest resources.The identification of tree species at the single tree scale provides more detailed tree species information for the detailed investigation of forest trees,and also lays the foundation for the extraction of forest single tree parameters and structural information.The airborne CCD image data is easy to acquire,less affected by the atmosphere,and has rich spatial structure information.It has unique advantages for single tree segmentation and tree species identification.At the same time,the airborne LiDAR technology has been widely used in forestry investigations due to its small limitations and strong penetration of the forest canopy.It has become one of the effective ways to accurately segment the forest canopy.Therefore,the combination of airborne CCD images and LiDAR data for single tree species classification has great potential.The traditional remote sensing tree species classification is limited by the data,and most of them are regional or stand-scale classification,which is difficult to realize the identification of single tree types.On the other hand,the classification methods are mainly based on pixel-based classification and object-oriented classification.However,for the high-resolution image classification based on the pixel-based classification method,it is easy to produce "salt and pepper" phenomenon,and both need to be screened for effective features.The deep learning technology can automatically extract the effective features of the image and input it with the original image to achieve end-to-end classification,……In this paper,Jiepai branch farm,Gaofeng national forest farm of Nanning City,Guangxi Province is selected as the research area due to its rich tree species..The airborne high-resolution CCD image and the LiDAR data acquired simultaneously are used as the data source to study the combination of deep learning classification and tree crown segmentation for single tree species identification.First,the U-Net and Res Net networks are combined to construct an improved Res-UNet network model.The convolutional layer of the U-Net network is represented by Res Net residual units.In upsampling,bilinear interpolation is used instead of inverse convolution,and Conditional Random Field(CRF)operation is added at the output of the network to optimize the model parameters through gradient back propagation.Then CRF post-processing was used to optimize the tree species classification map,and compared with the classification results using U-Net network and Res Net network alone to verify the classification performance of the proposed network.The improved Res-UNet network can extract multi-scale and deeper tree species information of the image when performing feature extraction,and at the same time avoid the problem of network degradation in the process of network deepening.For the LiDAR point cloud data,the watershed segmentation algorithm based on the canopy height model and the distance discriminant clustering method based on the point cloud are compared to determine the optimal segmentation method for the single tree canopy segmentation.Finally,the obtained segmentation results are combined with the improved Res-UNet classification results to realize the classification of single tree species.The improved Res-UNet network is used to classify the images in the study area.The overall and average accuracy is 90.03% and 88.39%,respectively,and the Kappa coefficient is 87.54%.Compared with the U-Net and Res Net networks alone,the classification accuracy has been improved by 22.90% and 13.45%,respectively.In single tree canopy segmentation,the watershed segmentation algorithm based on 0.8m resolution achieved the best single tree canopy segmentation accuracy,with a producer accuracy of 73.83%,a user accuracy of 80.05%,and an F-measure of 76.81%.Based on the improved Res-UNet network's tree species classification results and the watershed segmentation algorithm's crown segmentation results,the classification accuracy of single tree species is 73.13%.The research results show that the improved Res-UNet network proposed in this paper can better extract the deep spatial and spectral features of the image.Conditional random field post-processing operations can reduce the mixed phenomenon of broad-leaved tree species,make the classification boundary clearer and smoother.Bilinear interpolation instead of deconvolution effectively reduces the model parameters that need to be trained and the complexity of the model.Combining the proposed improved Res-UNet network classification algorithm with the single tree canopy segmentation based on the watershed segmentation algorithm can effectively realize the classification of single tree species in complex forest stands,and provide new ideas for the classification of single tree species.