Study On Automatic Extraction Of Urban Area Buildings Based On Convolution Neural Network

The low-earth orbit satellite has many sensors of different resolutions and different detection types,and provides researchers with a large number of satellite remote sensing images with abundant surface information.As one of the most important components of surface information(about 80% of the city's surface information),buildings have been extensively studied and applied in the fields of urban map mapping,urban infrastructure design and planning,land use coverage type survey and 3d digital city construction.It is one of that important content of satellite remote sensing image study that the characteristic information of building is extracted rapidly,accurately and intelligently in the remote sensing image of the satellite.In this paper,by adopting the idea of migration of deep learning and learning,improve traditional convolution neural network all FCN(Fully Convolutional Networks),the introduction of new Structure-cascading(Cascade Structure),and build a new depth of convolution neural network Structure,and use the building data sets to verify this algorithm.The main contents and conclusions are as follows:1)The architecture design and improvement of cascade full convolutional neural network: in view of the characteristics of local sensory field,low output resolution,multiple image loss,and fuzzy edge details of the traditional full convolution network,this paper introduces an empty convolution to increase the local sensory field of the feature image,and adopts the block method to ensure the output result of the same resolution as the input image,and the cascade network structure is introduced to maximize the information flow within the network.2)This paper studies the automatic extraction of buildings based on cascaded full convolution neural network: training the cascaded full convolution neural network proposed in this paper to automatically extract buildings and study the influence of the internal parameters of neural network and the super-parameters of neural network on the precision of the model.It is shown from that experiment result on the building data set of Massachusetts in the united state that the overall prediction accuracy of the proposed method is up to 92.3% compared with other methods,and the prediction accuracy of the two depth algorithm(76.3%?80.8%)is increased by 16% and 11.5% respectively.After obtaining the output result of neural network,the full-connection conditional randomfield algorithm is introduced and the output result is processed by the algorithm,and the result is smoother than the original result.3)Image classification and multi-objective segmentation application based on cascaded full convolution neural network:The internal structure of the cascaded neural network used for the second classification is adjusted,and the idea of transfer learning is adopted to replace the fixed nuclear convolution with multi-kernel convolution,and then the characteristic images of different scales are combined with data.The prediction accuracy of the proposed neural network model in UC Merced Land Use data set of the U.S.geological survey is over 99%,and the prediction accuracy of the International Society for Photogrammetry and Remote Sensing Vaihingen 2D Semantic Labeling dataset is 88.3%,exceeds the traditional neural network algorithm.