Point Cloud Classification Fusing Spectral Information Based On Deep Learning And Enhanced Post-processing

Point cloud classification is a very important link in point cloud data application.Applications such as city modeling,intelligent driving and indoor positioning all require high-precision classification results.In the process of classification,only using three-dimensional coordinates of point cloud to classify can only extract its spatial features,and it is easy to confuse objects with other objects which has similar structures.Spectral information is easy to obtain and data fusion is convenient.Adding spectral information can increase the distinction of various objects,effectively reduce the confusion of ground objects and improve the classification effect.Deep learning shows high performance in point cloud classification task,with characteristics of end-to-end output and powerful information mining ability,while point cloud data is disordered and irregular,and the amount of data is huge.Deep learning based on point cloud can dig deep spatial features in point cloud.Therefore,this paper uses spectral information and point cloud data fusion,and proposes a point cloud classification method fusing spectral information based on Point Net++ model.The research results are as follows:(1)Fusion method of spectral information and point cloud data.Using Vaihingen point cloud data set as the initial data,four types of fusion data sets containing different feature information are made by expanding the feature dimension of point cloud data.The method of data fusion is consistent with the method of neural network input data feature expansion.(2)Study the influence of spectrum and laser intensity information on point cloud classification based on deep learning.The network model based on Point Net++is used for comparative experiments,model training,prediction and accuracy evaluation respectively,so as to analyze the contribution of laser intensity information and spectral information in classification.The results show that the fusion of laser intensity and spectral information has the best classification effect,and the overall classification accuracy is 85.28%,7.68% higher than the initial data set,and the Kappa coefficient is 0.0994 higher than the initial data set,indicating that the fusion of spectral information can effectively improve the classification effect.(3)The classification enhancement method based on top-down clustering can effectively improve the classification accuracy.In view of the fact that point cloud classification with integrated spectral information is easy to cause the confusion of several vegetation categories,this paper proposes a segmentation method based on top-down integrated segmentation.Shrub and tree categories in the initial judgment results of neural network are separated by individual trees,and then the remaining low points are segmapped by DBSCAN clustering method based on density.After the segmentation,the normalized height and geometric neighborhood characteristics of the clustering point cloud were extracted to make the clustering point cloud data set.Random forest model is used for training and final prediction to achieve the purpose of classification enhancement.The experimental results showed that the overall classification accuracy of the enhanced results was improved by 1.5%,the Kappa coefficient was improved by 0.0189,47.2% of the grassland prediction was corrected in the clustered point cloud,and the F1 score of tree classification was improved by2.7% and shrub classification was improved by 3.2%.It shows that the clustering method has better effect on vegetation classification and can strengthen the effect of point cloud classification.Compared with the five methods provided by ISPRS website,it scored the highest in roof and tree category F1 score comparison and achieved the best classification effect.