Research And Application Of Semantic Segmentation Of High-Resolution Remote Sensing Image Based On Deep Learning

High-resolution remote sensing image is one of many remote sensing image data types.It is divided into aerospace and aerial remote sensing images according to different shooting heights.It has rich information of geometric features and attribute details,and its visual effects are intuitive.It is widely used in military and civilian fields.In the application of environmental assessment,disaster assessment,forestry surveying,precision agriculture,urban planning,map production and updating,change detection,and military target recognition,it is necessary to use the method of visual recognition for high-resolution remote sensing images to analyze the scenes.Semantic segmentation is one of the ways to analyze image scenes.It can simultaneously perform image pixel classification and image object segmentation by labeling each pixel of the image with semantic tags.Due to the inherent phenomena such as"different spectra of the same kind of objects"and"similar spectra of different objects"on high-resolution remote sensing images,this is the main reason for the low precision when using traditional methods for semantic segmentation.In recent years,image semantic segmentation is mainly realized by deep learning.However,there are not many researches and applications of deep learning in high-resolution remote sensing images.Therefore,this paper is based on the related algorithms of deep learning and deeply studies the semantic segmentation of high-resolution remote sensing images.The main work is as follows:(1)Establish a high-resolution remote sensing image semantic segmentation dataset:Aiming at the time-series image semantic segmentation task,this paper builds a remote sensing image semantic segmentation dataset with new buildings as semantic tags.At the same time,the related processes of establishing remote sensing image dataset are introduced in detail,including dataset acquisition,manual annotation,image preprocessing and dataset partitioning.Finally,six data augments were made for the three original datasets used in this experiment.(2)Solve the problem of low semantic precision of single remote sensing image:By analyzing the FCN-based image semantic segmentation method,it is found that the simple splicing fusion operation used does not greatly improve the segmentation precision.On this basis,this paper proposes a DenseFPN based on DenseNet's feature pyramid semantic segmentation network.The network has three advantages:First,the feature extraction part is completed by using the DenseNet structure,and the feature extraction capability of the network is enhanced by means of dense connection;Second,the fusion part is completed by using the feature pyramid structure.The optimized unit of the design makes the utilization degree of each level feature the same,and then the splicing unit is used to fuse the feature of each level,thereby enhancing the spatial information recovery capability of the network;Third,it uses the transfer model to improve the performance of the network.In this paper,the DenseFPN network is experimentally optimized and verified on the urban remote sensing image dataset and remote sensing image road extraction dataset,respectively achieving 82.2%and 86.8%accuracy,which is 14.0%and 15.4%higher than the FCN accuracy.In addition,the network performed the best on small-sized targets,especially the semantic segmentation of cars increased by 30.9%.(3)For the semantic segmentation of time-series images,it solves the problem that the traditional scheme steps are separated,the labeling amount is large,and the result noise is large:This paper designs a time-series image semantic segmentation network DAUnet,which is an end-to-end framework with a small amount of annotation and no noise caused by the difference.DAUnet needs to input time-series images in a spliced form.The network mainly includes three parts:the encoding process,the multi?scale fusion process and the decoding process.It has made the following three improvements:First,the combination of DenseNet and dilated convolution improves the feature extraction capability of the network and maintains the resolution of the feature map unchanged;Second,enrich the diversity of feature fusion by paralleling more cells in ASPP;Third,using 1×1 convolution in the Unet decoder reduces the number of feature maps,thereby reducing the amount of computation.Based on the self-built time-series building extraction dataset,the DAUnet semantic segmentation network is used to identify the newly added buildings and obtain their outlines.The IoU of the newly extracted buildings is 79.5%,the three improved precisions are respectively increased by 14.6%,2%,and 0.7%.