LiDAR Data Urban Terrain Classification Based On Deep Learning

Light Detection and Ranging(LiDAR)data records the height of different objects on the ground and provides rich elevation information,so it plays an important role in the terrain classification.In recent years,deep learning models have been widely used in terrain classification of LiDAR data due to their strong feature extraction ability.However,these models usually require a large number of training samples to obtain good results.But in the field of remote sensing,data shortage is a common phenomenon,so it limits the development of deep learning models in LiDAR data classification.In addition,most existing models are designed by hand for the characteristics of one data,but lack universality for other data.The LiDAR data captured by different sensors have different properties,so the corresponding models should be designed automatically for different LiDAR data.This paper puts forward two solutions to the above problems respectively,and the main research contents include the following parts:Firstly,the theory of generative adversarial network and its derived model are deeply analyzed,and the theory of neural architecture search is studied,which laid a theoretical foundation for the construction of LiDAR data classification model in the following chapters of this paper.Secondly,the structure of conditional generative adversarial network is improved.Residual unit is introduced to increase the network depth of generator to improve the quality of generated samples,and the generated samples are extended to the training data to fully train the model.The structure of discriminator is improved so that it can output multiple categories,and the Drop Block regularization method is introduced to improve the classification ability of the model.The experimental results show that the proposed method significantly improves the classification accuracy of LiDAR data and greatly alleviates the problem of data shortage.Finally,a LiDAR data classification method based on neural architecture search is proposed.A new search space is designed and attention mechanism is introduced to improve the feature extraction capability of the model.The cosine power annealing learning rate adjustment strategy is introduced to reduce the time consumed in search and evaluation stage.The label smoothing regularization method is introduced to enhance the robustness of the model and avoid the over-fitting phenomenon.Experimental results on three LiDAR data obtained from different sensors show that the proposed method achieves the best classification performance.