Automatic Recognition Of Potential Hazardous Buildings Along The High-speed Railway In The Framework Of Scene Interpretation

The geographical environment is very complex and there are many potential hazard along the high-speed railway(HSR).Specifically,houses,factories and other illegal buildings along HSR is harmful for the safety of HSR.Consequently,it is necessary to investigate the potential hazardous buildings along HSR.Field investigation is used to detect potential hazardous buildings generally,it is costly and time-consuming,this method is difficult to monitor statuses of all the line network effectively.High resolution remote sensing has many advantages such as timeliness and periodicity,it can be used to detect the potential hazardous buildings rapidly,objectively and dynamically.Details of objects are abundant in the high resolution remote sensing image,however,because of same objects with different spectrums,different objects with same spectrums,there is a loss of building extraction accuracy with pixel-based method.The accuracy is improved due to the object-oriented segmentation considering the spatial relationship among pixels,however it is difficult to determine the optimal segmentation scale.In addition,the two methods detect the buildings based on low-level visual features rather than higher-level semantic features of images,so there is a clear semantic gap which affects the accuracy of buildings extraction.So,it is necessary to understand images from higher level of scenes.A Google image of Suzhou-Bengbu in Beijing-Shanghai high-speed railway is used in this study.In the framework of interpretation of scenes,we create sample database firstly,then we divide images into overlapping image blocks.Image blocks are regarded as the documents,visual words histogram and mixed ratio of potential semantic topic are achieved through bag of visual words model and the Latent Dirichlet Allocation model respectively.Categories of image blocks are achieved with SVM classifier,category of each pixel is determined by category voting,potential hazardous buildings are recognized automatically;image blocks are input into the trained convolutional neural network,full connection layers are created through convolution,pooling and full connection,the probability of categories of image blocks are achieved through the Softmax classifier,the probability of category of each pixel is achieved by the equal weight average,we choose the category whose probability is maximum as the category of the pixel finally,potential hazardous buildings are recognized automatically.The conclusions are as follows:(1)The methods in the framework of scene interpretation can improve compactness and completeness of results compared to pixel-based and object-oriented methods that use low-level features,the overall accuracy and producer accuracy reach 91%,the kappa coefficient reaches 0.71,the results are close to the ground.(2)Convolutional neural network avoids the limitation and blindness of artificial features used in bag of visual words model and topic model among methods in the framework of scene interpretation,it has best performance results.