Research On Extraction Method Of "Two Violation" Buildings Based On Deep Learning

Buildings are an important part of urban planning and management,but with the continuous improvement of China's urbanization level and the strengthening of urban appearance management,the phenomenon of "two illegal" buildings with illegal land use and illegal buildings as the main body has become an important factor affecting economic development,social stability and people's livelihood.Therefore,the efficient detection of "two violation" buildings and the effective supervision of "two violation" business are one of the keys to ensure social stability,safeguard civil rights and improve residents' happiness.The research of "two violation" building extraction method is helpful to improve the utilization rate of natural resource data and promote the construction of urban and rural information government.At the same time,as the frontier application of remote sensing and UAV,it has very important research significance and broad application prospects.Traditional extraction methods based on object segmentation and building features need more artificial experience to design feature extraction methods.Building extraction based on deep learning has strong self-learning ability,and with the maturity of theory and hardware,it is gradually taking the place of traditional extraction methods,and has become a hot spot of current research.The deep convolution neural network is used to extract the target features layer by layer,and the potential high-dimensional feature information is extracted in the form of dimension elevation,which can effectively reduce the image differences caused by the external shallow features of remote sensing image,such as color,tilt and so on,so as to improve the final building extraction accuracy.On the basis of building extraction contour,through the change detection and rule base and the final business system,on the basis of building extraction results,it realizes the judgment of suspected "two violation" buildings and the supervision of "two violation" business,realizes the "two violation" closed loop,and improves the work efficiency.Therefore,this paper focuses on building extraction algorithm and illegal interpretation based on deep learning.(1)Building data set construction and expansion in Nan'an areaBuilding extraction model training based on deep learning is inseparable from high-quality and large number of data sets.At present,whu data set published by Wuhan University in 2019 is the most comprehensive data set for building extraction in the world,with complete types,large coverage area and high resolution.However,there are still many differences in building style,density and color between it and the data source in this paper.Therefore,this paper analyzes the current situation of regional data,using manual annotation,combined with brightness transformation,oblique stretching and other data augmentation methods,effectively solves the problems of missing data sets,different angles and brightness of multi-phase remote sensing images,meets the actual training requirements of the model,and makes it suitable for the actual business background.(2)Automatic building extraction algorithm based on deep learningThe automatic extraction of buildings based on deep learning is the key to effectively reduce the manual workload,and it is also the basis of subsequent suspected "two violations" building interpretation.Based on the existing data sets,this paper compares several building extraction algorithms,and selects mask r-cnn algorithm as the extraction algorithm.Combined with the actual business needs,on the basis of the original algorithm,the parameters are adjusted to sacrifice part of its precision to ensure its high recall rate,so as to meet the application requirements.On the basis of the traditional verification method,the accuracy and recall rate under four different conditions in the business background are verified,the output is given,and the input of the subsequent suspected "two violations" pattern judgment is explained.(3)Judgment of suspected "two violations"The judgment of suspected "two violations" includes two steps: extraction result change detection and suspected violation judgment.First of all,change detection is to preprocess the model output data on the basis of building extraction results,study the specific methods and steps of change detection,and detect three types of change graphs of new,reduced and comprehensive changes for subsequent suspected illegal interpretation;secondly,combined with policy documents and actual business,use the rule engine to construct the rule base of existing rules In addition,it also enriches the rules of human experience to assist interpretation.Finally,through the comparative analysis of the suspected "two violations" results extracted by this method and the actual "two violations" results after field verification,the final accuracy verification results are given,the accuracy reaches 90.6%,and the recall rate reaches 89.2%,which basically meets the actual business needs.(4)Comprehensive rectification platform for "two violations" monitoringBased on the deep learning "two violation" building extraction method,this paper extracts the suspected "two violation" pattern,combined with the actual business processing needs,constructs the "two violation" monitoring and comprehensive rectification platform,from building extraction,suspected "two violation" pattern judgment,case allocation to illegal audit,and finally by the suspected confirmed as the actual "two violation" pattern,forms a complete "two violation" pattern In order to provide strong support for improving the efficiency of government work and strengthening supervision.