| Extracting artificial buildings from remote sensing images,such as roads and buildings,plays an important role in many urban applications,such as urban land use and land cover assessment,updating geodatabases,change detection,and so on.This task is often difficult due to complex data with heterogeneous appearance forms,large in-class and low-class-to-class variations.The fitting of building boundaries has always been an extremely important research direction,and in practical applications often requires more aesthetic building boundaries to meet the needs of the supplier industry.Beijing No.2 remote sensing imagery combined with its specific index data NDVI and BAI is used in this thesis,which is based on n DSM.The study used the integrated algorithm classifier of support vector machine,Bagging,and Boosting to complete the extraction of buildings.Finally,the extracted building boundaries are enhanced to a certain extent,and the pix2 pix network model is used for training and inference on the Google Colab platform.The findings suggest:(1)Using Python language,n DSM data obtained based on laser point cloud data,and auxiliary other exponential data,the accuracy of building extraction is improved,and the shortcomings of insufficient feature data of a single remote sensing data source provide technical support for the extraction of buildings and the extraction of other land types.(2)After adding auxiliary data,the cell-oriented method is used to extract,and the algorithm of machine learning is selected compared with deep learning,which requires only CPU,and the training time is shorter when the accuracy meets the requirements.The algorithm of the integrated class better reduces noise and improves the accuracy of the algorithm.(3)Based on the pix2 pix network model,the network model can successfully fit the boundary line of the building,the method has the advantage of end-to-end,can successfully connect the disconnected boundary line,the frizzy boundary line,etc. |
PDF Full Text Request