High-resolution Remote-sensing Image Buildings Extraction Methods Based On Deep Learning

Building extraction is an important content of geodatabase updating and construction,and building extraction using high-resolution remote sensing images is an important direction of this research and an important content of remote sensing frontier technology research.With the accelerated urbanization and rapid development of remote sensing technology,urban remote sensing has entered the era of big data,and the corresponding deep learning method has become an important tool and a popular research field for building extraction.The study of building extraction in remote sensing images using convolutional neural network method has made some achievements,but further research is needed to explore better extraction methods in terms of building extraction accuracy and network training speed.In this dissertation,a systematic study of deep learning techniques for building extraction from high-resolution remote sensing images is conducted,and the main work is as follows.(1)Aiming at the multiscale characteristics of buildings with various shapes and irregularities,and the problems of incomplete extraction of large-sized buildings and missed and wrong detection of small-sized buildings that may occur when using convolutional neural network algorithm to extract buildings,an improved method of extracting buildings based on ERFNet network for high-resolution remote sensing is proposed,in which the depth separable convolution is introduced into the residual block of bottleneck structure in the coding layer part.In the coding layer part,the depth separable convolution is introduced into the residual block of the bottleneck structure and the Channel Shuffle structure is used to realize the information exchange between groups;the depth convolution and coordinate attention mechanism are used in the downsampling block to enable the model to locate and identify the target more accurately;the CARAFE upsampling operator is added in the decoder part to improve the feature extraction ability of the network.The improved extraction method has higher semantic segmentation accuracy and better recovers the spatial details of buildings.(2)Aiming at the problems that the Deep Labv3+ model is not accurate enough to segment the edges of remote sensing image buildings and the number of network parameters is large,a lightweight Deep Labv3+ model for remote sensing building extraction is proposed,using the lightweight network Mobile Ne Xt to replace the backbone network Xception of Deep Labv3+ to reduce the number of network parameters and improve the training speed.The connection method of Dense ASPP structure is adopted to perform stitching operation for feature layers with different scale information;the convolutional block attention module CBAM is introduced to perform feature mapping of local information on channel attention and spatial attention to achieve high precision building fast extraction,and the improved network is more complete for building edge extraction.The improved network is evaluated on two public datasets WHU building dataset and Massachusetts building dataset.The evaluation is based on the commonly used semantic segmentation evaluation metrics of remote sensing images to evaluate the accuracy and precision of the results more comprehensively and objectively.The experimental results show that the two improved building extraction methods in this dissertation can be better applied to the extraction of buildings in high-resolution remote sensing images.