Research On The Classification Of Multispectral LiDAR Point Cloud Data Based On Deep Learning

Land cover is becoming more and more important for land resource management,ecosystem protection,urban planning,and sustainable development,etc.Using remote sensing data to classify land cover is an effective method to quantify land resources and monitor their changes.Multispectral light detection and ranging(Li DAR),as a new kind of active remote sensing technology,can simultaneously capture geometry and spectral information of objects,which has become a fast data acquisition for large scale areas.Multispectral LIDAR provides a new data source for environmental modeling,disaster response,and land cover classification,etc.However,the characteristics(e.g.,disordered,massive,etc.)of multispectral LIDAR point cloud data bring challenges to effectively extract land cover features.The traditional threedimensional(3D)point cloud processing algorithms achieved poor performance in semantic features representation due to point density variation and distributions.With the increasing development of deep learning,deep learning has been widely used in 3D point cloud land cover classification and target detection.Thus,it becomes a important research topic how to use deep learning to enhance abstract description of target features in multispectral LIDAR point cloud data and improve land cover classification accuracies.Therefore,we propose a deep learning-based land cover classification method using multispectral Li DAR point cloud data,which includes multi-wavelength data fusion,training sample generation,and 3D point cloud classification of Point Net++ combined with a channel attention mechanism.This research can improve the levels of intelligent interpretation of multispectral Li DAR data and the automation degrees of target recognition.This research mainly includes the following specific contents:(1)Multispectral Li DAR data preprocessing.The pre-processing study includes multi-band data fusion,data labeling,and full-coverage training sample generation,which provides an effective data input for land cover classification of multispectral Li DAR data.(2)We propose a SE-Point Net++ model for multispectral Li DAR point cloud land cover classification.To address high computational cost and low classification accuracy caused by a useless channel feature learning process,we build the SE-Point Net++ model,a combination of a channel attention mechanism(Squeeze-and-Excitation block,SE-block)and the Point Net++model,which improves the overall performance of the proposed network architecture.Furthermore,because the proposed module considering the geometric structures complexity of objects and the spectral reflection differences of ground features,the proposed module obtained a more robust expression effect of the ground features,and further achieved the purposes of efficient and intelligent interpretation.(3)Land cover classification experiment of multispectral LIDAR point cloud data based on SE-Point Net ++.Using the two datasets in different regions collected by a Optech Titan airborne multispectral LIDAR system,we discussed the influence of input data types and sampling scale strategies on the accuracy of land cover classification,and verified the superiority of SE-Point Net ++ model in multispectral Li DAR data classification by comparing with the classical deep learning model.