| With the continuous development of airborne LiDAR(Light Detection and Ranging)system and the expansion of industry applications,the theory of airborne LiDAR point cloud data processing has been widely studied.Among them,point cloud classification is a necessary step for various 3D scene understanding and applications in various industries,such as urban3 D modeling,power line inspection,and terrain mapping.Traditional point cloud classification methods not only fail to extract deep-level features due to the reliance on manual features and fixed machine learning algorithms,but also have little automation,many parameters and complex processes.While the deep learning has the powerful ability to express in-depth information and to automatically extract high-level features for classification.In view of the lack of effective screening of a large number of input features in the study of airborne point cloud classification methods,and the existence of large loss of point cloud features and insufficient learning of point cloud features in the classification process,that makes the deep learning model poor learning efficiency and low classification accuracy.Based on the deep learning point cloud classification algorithm Rand LA-Net network,this paper presents an indepth study on the improvement of the Rand LA-Net network for proximity scene classification based on the characteristics of airborne point cloud data and the fusion mode of airborne point cloud features on the input side of the network.The main work and innovation points of the paper are as follows:1.The current status of airborne LiDAR point cloud classification research is summarized,the shortcomings of existing classification algorithms are analyzed,and the characteristics of airborne LiDAR point cloud data are outlined.It is also indicated that methods to improve airborne point cloud classification are studied from the perspectives of deep learning network models and point cloud features at the input of the network.2.Proposed a feature filtering algorithm based on random forest to construct fewer and better airborne point cloud feature dataset.By analyzing and summarizing various point cloud features,a multi-scale and multi-featured point cloud feature dataset is established,and a random forest approach is proposed to obtain the optimal point cloud feature,which reduces the feature dimensionality of the point cloud and the input of irrelevant features;Meanwhile,by aligning and fusing the airborne point cloud with the spectral image,the airborne point cloud data with spectral information is obtained.3.Four airborne LiDAR point cloud feature fusion classification models were constructed and the near-field point cloud classification network Rand LA-Net was improved.To validate the Optimal Feature Fusion-Based Spectral Information Network(OFFS-Net)proposed in this paper,a total of four point cloud feature fusion classification models were constructed for experimental comparison and analysis.By improving the Rand LA-Net point cloud classification network to enhance the adaptability to large area and low density airborne LiDAR point clouds,retaining the advantage of random sampling in the Rand LA-Net network,and merging the constructed point cloud features with the local feature fusion module in the network,which enables the network to learn deeper semantic information.4.In order to verify the robustness and generalization of the OFFS-Net model algorithm proposed in this paper,two publicly available airborne LiDAR point cloud experimental datasets,the Vaihingen and the LASDU,were used for experiments.In terms of the comprehensive evaluation metrics OA and Avg.F1 of point cloud classification results,the OFFS-Net model achieves 84.9% and 72.3% on the Vaihingen dataset;88.0% and 79.7% on the LASDU dataset,respectively.Compared with other point cloud classification model algorithms,the model proposed in this paper has the best classification results.It further demonstrates the effectiveness of the proposed OFFS-Net point cloud classification method based on the advantage of fusing geometric feature and spectral information. |
