Machine Learning For Cloud Extraction From Satellite Imagery And Filtering Of 3D Point Cloud

With the continuous progress of remote sensing sensor technology,the amount of remote sensing data is rapidly expanding,but the corresponding technology of information extraction from remote sensing data has not made a breakthrough.Although scholars have done a lot of research on automatic information extraction algorithms,due to the complexity of the problem,many of the achievements still remain in the experimental stage,or need a lot of manual labor support,not only consume a lot of manpower and material resources,but also pull down the production efficiency,the face of the huge amount of remote sensing data has become a serious shortage.Based on the above considerations,both scientific research and production have an urgent demand for technology of information extraction from remote sensing data with high precision and high degree of automation.This paper aims for the method of information extraction from remote sensing data with high precision and high degree of automation,exploring the methods of automatic cloud extraction from satellite imagery and automatic filtering of airborne laser point cloud.The main contents of this paper include:1)A method of automatic cloud extraction from satellite imagery is proposed.Aiming at the large morphological and size changes of cloud,this paper describes the images of cloud using the BOW model,which is widely used in text processing and image processing.The model uses the clustering centers clustered by the features extracted from the training samples in the feature space as the visual words,and then generates robust descriptors for images based on the response histograms of these visual words from the dense local features extracted from the images.Then the descriptor are used to train and classify images.Finally,the GrabCut algorithm is used to further optimize the classification results.2)A method of three-dimensional laser point cloud feature mapping is proposed.For each laser point in the three-dimensional space,the relative position information of the adjacent points within a certain range is extracted and mapped,so that the features extracted for each point are mapped into a two-dimensional feature map to facilitate the subsequent training and classification.3)A method of three-dimensional laser point cloud automatic filtering is proposed.After extracting the information from the massive training sample points and mapping them into two-dimensional feature maps,a well-constructed deep convolution neural network is trained and then used to classify the unclassified points.4)A method of result refinement for three-dimensional laser point cloud filtering is proposed.It's not enough to classify each point isolated for laser point cloud filtering,which may cause the triangulation obtaining obvious burr due to individual errors even the error rate is low in some cases.Considering the goal of point cloud filtering is to extract relatively precise and smooth surface of earth,this paper uses the full-connection conditional random field(CRF)to add a smoothness constraint based on features and labels of adjacent points to further optimize the filtering results of convolution neural network to make the final results more in line with production requirements.