Deep Learning Based Remote Sensing Image Building Extraction And Change Detection Research

With the rise of remote sensing technology,a large number of high resolution remote sensing images have been produced.Although the technology has matured to the extent that high-resolution remote sensing images can clearly identify buildings,roads and other information,the dense and relatively small targets in the images would take a lot of time and effort to analyse and interpret these images using only manual visual inspection.At the same time,as urbanisation grows,buildings are an important element and it is important to keep abreast of changes in buildings for urban planning.Therefore,it is an important research direction to automatically analyse the information and changes of buildings in a large number of remote sensing images.In recent years,deep learning technology,with its powerful feature learning capability,has been widely used in the fields of speech recognition,image recognition,natural language processing and so on.Compared to traditional feature acquisition methods,deep learning algorithms have the advantage of automatically extracting features at all levels,which is beneficial for obtaining more information about the features.Therefore,deep learning-based methods provide a new way to extract and analyse the distribution and changes of buildings in remote sensing images.In this thesis,Unet and Unet++ are selected as the base models for optimisation to address the problems of mis-detection,missed detection and edge blurring that occur in the current algorithms in the process of building extraction and detection,and are used for building recognition in a single image and building change detection in two time-series images respectively.The main research elements are as follows:(1)Aiming at the problem of complex types of remotely sensed images and the difficulty in acquiring them,this thesis provides a basic introduction to remotely sensed images and summarises the datasets related to buildings and their change detection to provide assistance to relevant researchers.(2)Aiming at the problems of small targets in remote sensing images and difficulty in distinguishing target edges,a building extraction model based on multi-level feature fusion is proposed.The model is improved on the basis of Unet in two directions,feature extraction and feature fusion.In the feature extraction part,the traditional convolution block is replaced by a multi-scale residual module,which extracts building features layer by layer and continuously fuses feature maps with different perceptual fields,thus obtaining correlations between different building features.Secondly,a parallel expansion module constructed from cavity convolution is introduced in the feature fusion section to fuse the features obtained in the downsampling process with the symmetric upsampling section after further fine extraction,reducing the semantic gap between codecs,while the obtained deep feature maps are further concatenated with the feature maps obtained in the shallow layer by means of a jump connection to achieve feature fusion,thus achieving accurate extraction of buildings.The result is accurate building extraction.Experiments on remotely sensed building datasets show that the model can better achieve building segmentation extraction.(3)In response to the problem that the single-branch network of early fusion cannot extract features separately according to the characteristics of each image,which easily causes information redundancy and confusion,a dual-branch remote sensing image change detection model is proposed.The feature extraction part of the model is designed based on Unet++ in the coding part,and the multi-scale convolution module designed in this thesis is still used in this part to replace the convolution module in the original Unet++ for change feature extraction.In the decoding part,the spatial and spectral features extracted through the weight-sharing twin network are combined with deep supervision,and the training process of different upsampling branches is supervised by the objective function,which in turn reduces blind training and improves the training purposefulness.In addition,in the depth-supervised part,an attention mechanism is introduced,which is mainly used to adjust the weights of the fused feature maps during decoding,aiming to suppress the interference of noisy information and to highlight the change part of the region of interest.Experiments on the remote sensing building change detection dataset show that the model can better identify the changing regions of buildings in high-resolution remote sensing images.This thesis has 38 figures,9 tables and 82 references.