Research On The Method Of Building Damage Assessment With High Resolution Remote Sensing Images

Building damage assessment is of great significance for earthquake emergency response and post-disaster reconstruction.The key to the problem is how to obtain reliable results with limited time and data.Because the data of pre-earthquake images and the corresponding geographic information systems are not always available in practical applications,it is more valuable to carry out research on building damage assessment methods with only post-earthquake image data.Based on the previous work,this paper studies the damage assessment methods of buildings in high-resolution remote sensing images after the earthquake.The main contents are as follows.Based on the characteristics of the remote sensing image data itself,this paper proposes to detect the interference of the building area in the low-resolution multi-spectral image to eliminate the background area,and then use the nearest neighbor diffusion algorithm to perform the full-color image of the output area ROI Fusion to improve the spatial resolution of the method.The method can improve the accuracy of building damage assessment,but also to ensure the detection efficiency.In the stage of building area extraction,an area extraction algorithm based on FCN network is proposed,which can not only achieve end-to-end output,but also achieve high efficiency by GPU acceleration.In order to solve the problem of large amount of remote sensing image data and less annotation data,a two-stage training method of parameter fine-tuning is studied,and a method of synthesizing large-scale annotation data based on GoogleMap and OSM data is adopted.With large-scale online data to improve the accuracy of the model,to solve the problem of insufficient data marked.According to the characteristics of the building area,the post-processing method based on mathematical morphology is studied.In the stage of building damage assessment,an object-oriented image classification method based on transfer learning is proposed.This method introduces the feature extraction method based on the migration learning of convolution neural network in the framework of the usual object-oriented image classification.This paper compares the similarities and differences of several transfer learning methods and gives the characteristics that are most suitable for object-oriented image classification Extraction method.Secondly,in order to solve the problem of parameter sensitivity of the algorithm,this paper proposes a way to integrate multiple window scale features.Experiments show that the improved method proposed in this paper can improve the classification efficiency and accuracy compared with the existing algorithms.