Extraction Of Buildings Along High-speed Railway From High Resolution Remote Sensing Images Based On Convolutional Neural Network

Buildings along the high-speed railway and accompanying human activities may cause problems such as invasion of foreign objects and settlement of roadbeds,which are poten-tial factors affecting the safety of high-speed rail operations.Therefore,regular monitoring of the buildings on both sides of the high-speed railway,timely detection and elimination of potential hazards,will help improve the safety of high-speed rail operations.At present,the monitoring of buildings along the high-speed railway is mainly carried out by manual periodic inspection.However,this way is time-consuming and labor-intensive,and some areas are even hard to reach,making it difficult to meet the needs of China's increasingly large high-speed railway network.High-resolution remote sensing imagery has the advantages of timeliness,comprehensiveness,and high economy,and the automatic information extraction technology has also become increasingly mature,which provides a new technical means for monitoring of buildings along high-speed railways.The premise of monitoring buildings is to extract the building information on the im-ages.Although buildings have obvious characteristics in images,their shapes,sizes,and tex-tures are complex and various,and high-resolution images are difficult to provide sufficient spectral information.Traditional building extraction algorithms are difficult to apply wildly.The high-speed railway has wide-area and strip-shaped space characteristics,and the build-ings along the line are extremely complicated,which not only further increases the difficulty of extracting buildings,but also poses a challenge to preprocess these remote sensing images.Covering the area along high-speed railway requires the connection of multiple high-resolution images.Image matching is a crucial step in connecting these images.However,differences in the image quality of these images,which will cause matching difficulties and affect the subsequent analysis.Learning the relationship between image quality and image matching will help to screen images in advance and provide empirical guidance for image matching.In recent years,deep learning represented by convolutional neural networks(CNN)has made breakthrough progress in computer vision,and has shown the potential of extracting buildings along high-speed railway.CNN needs a large number of sample data for training.There are two kinds of samples,e.g.,scene-level and pixel-level samples.The scene-level samples are easy to obtain and can be used to quickly extract buildings,but the accuracy is low; while the pixel-level samples are difficult to obtain,but the accuracy is high.This thesis focuses on the characteristics of buildings along high-speed railways and bases on the CNN to study the extraction methods of buildings along high-speed railways with the scene-level samples and pixel-level samples.This thesis studies the image matching from two aspects,i.e.,feature extraction and fea-ture matching.At first,based on analysis,the peak signal-to-noise ratio and structural similar-ity index are selected as the measures of image quality; feature repeatability and correctness rate are selected as the measures of the performance of feature extraction and feature matching,respectively.Then the selected 9 images along high-speed railway with different complexity are processed with different degrees of blurring and noise addition to obtaining a series of im-ages of different quality.At last,the matching experiments are performed on these images to obtain the relationship between the image quality and image matching.For feature extraction,the results show that there is a quantitative relationship between image quality measure and feature point repeatability.When the scene and degrading factors are determined,this quan-titative relationship is monotonous and can be easily simulated with some simple functions.For feature matching,the above conclusion can also be obtained.In addition,the relationship between the resolution difference and feature matching is also studied in this thesis.The ex-perimental results show that when the ratio of matching images' resolution is less than two,these images can be matched well.According to the characteristics of the building along high-speed railway,this thesis pro-poses a strategy of integration of the multilevel and multiscale feature fusion.Under that strategy,two fully convolutional neural networks are proposed,i.e.,mc Res Net-cs AM and DPWRes Net-FUM.mc Res Net-cs AM consists of a multiconnection residual network and a class-specific attention model.The former fuses multilevel deep features without introduc-ing any redundant information from low-level features.The latter can learn the contributions from different features at each scale.DPWRes Net-FUM utilizes the proposed dilated pyra-mid wide Res Net to extract and fuse multiscale features at different levels and uses a fine upsampling module to restore the localization of the features.Extensive experiments on the proposed pixel-level building database indicate that the two proposed networks outperform other state-of-the-art methods with 1 achieving 85% and 87%,respectively.It is worth men-tioning that mc Res Net-cs AM also achieves the highest overall accuracy of over 50 methods in the online test of the ISPRS Potsdam dataset,with the highest 1 for building and impervious surface.For the proposed methods,DPWRes Net-FUM performs better than mc Res Net-cs AM in two building datasets.In terms of complexity,DPWRes Net-FUM is more efficient than mc Res Net-cs AM and the model size is similar.With only scene-level samples,this thesis proposes two weakly supervised methods for extracting pixel-level building information based on CNN.The first method utilizes the heat map generated by CNN under the mechanism of probability overlap.The experimental results show that this method is better than other scene parsing methods under that mechanism,but the results are relatively rough and it is only suitable for extracting the region of buildings along high-speed railway.The second method is to train the fully convolutional network by the pseudo pixel-level labeling map generated by the heat map.According to the characteristics of buildings along high-speed railway,the hierarchical fine optimization and the scale self-supervised optimization are proposed to improve the detail and homogeneity of the buildings on the heat map.The results of experiments indicate that this method obtains more precise results than the first method.For the high-speed railway building dataset,this method reaches the highest 1(72.8%)of all,and its results are better than other methods in visual comparison.