Research On The Method Of Extracting The Building Seismic Damage Information Based On Improved U-Net

Earthquake disaster is one of the most serious natural disasters that threaten human survival.After the earthquake,the timely acquisition of building seismic damage information is of great practical significance for post-disaster assessment and emergency rescue.Remote sensing technology has the characteristics of fast,large-scale and all-day time.It provides large-scale macroscopic image data in time after an earthquake disaster,and provides strong data support for the extraction of earthquake damage information of buildings.Among the traditional methods for extracting building seismic damage information,the most commonly used methods are pixel-based and object-oriented methods,the image features extracted by the pixel-based method are too simple,there is difficult to apply to post-earthquake remote sensing images with complex ground background.The optimal segmentation parameter is difficult to determine in the object-oriented method,which restricts the accuracy of information extraction.In recent years,deep learning technology has developed rapidly.This method can learn more advanced features from a small amount of preprocessed or unprocessed data,and has achieved application effects that are difficult to achieve by traditional algorithms in the field of image segmentation.This thesis applied deep learning methods to extract seismic damage information of buildings in earthquake disaster scenarios and combined remote sensing technology with cutting-edge computer technology in order to improve the accuracy and precision of building seismic information.The main research contents and results of this thesis are as follows:(1)Aiming at the problem of lack of sample data sets for building damage after earthquakes for model training,comprehensively analyzed the interpretation characteristics of building damage information on remote sensing images.Based on aerial image data in the study area,a small sample size data set was established for model training and verification through labeling,random selection,and data enhancement.The seismic damage information of buildings in the data set was divided into three categories: basic intact,damaged,and completely damaged according to the degree of damage.(2)Aiming at the problem that the depth of the U-Net network is slightly insufficient,it is impossible to clearly characterize the complex features and affect the accuracy of building seismic information extraction.Combined residual thought and dilated convolution technology to construct efficient residual dilated convolution block to enhance the ability of model feature extraction.Improved and optimized the U-Net network architecture through regularization technology to construct an E-Unet(Efficient U-Net)model.(3)In order to verify the validity of this article,the E-Unet network and U-Net,Seg Net network were qualitatively and quantitatively analyzed on the same test dataset.The results show that the E-Unet network model constructed in this thesis has a higher accuracy in extracting seismic damage information of buildings.The average overall classification accuracy on the test data is 93.2%,the kappa coefficient is 0.8232,and macro-P and macro-F1 are 88.26%,86.55%.E-Unet has a stronger ability to distinguish different types of earthquake damage information,and the average extraction accuracy of the three types of basic intact,damaged,and completely damaged are 89.73%,80.39%,and 87.59%,respectively.