| With the development of spatial information technology,the number of high-resolution remote sensing images has increased rapidly.Because of its wide coverage of the ground and the characteristics of timely updating and monitoring the ground situation,it has application value in digital city,disaster assessment,resource monitoring and so on.Previously,researchers mainly relied on manual methods to extract image information,which was time-consuming and labor-consuming.In order to analyze image information more efficiently,the deep learning developed in recent years has been gradually applied in the extraction of remote sensing image information by scholars,with initial results.As the main component of urban area,buildings have important reference value to urban research.Therefore,this paper realizes automatic extraction of buildings based on deep learning related algorithms,and selects UNet as the basic algorithm for optimization,mainly completing the following work:(1)Create data sets and preprocess it.The data sets include two parts:the image data taken by Gaofen No.2 and the artificially annotated truth map.Production of experimental data:first,expand the effective sample set by horizontal flipping,vertical flipping,and rotation to enhance the invariance of objects and make the limited data produce more value;second,adjust the color saturation of the image to make the image content more clear;third,before cropping,fill 256 size around the image,and use 256 overlap to crop the image,avoiding the problem of inaccurate edge prediction of the original image and 512×512 small-size image.(2)The experiment compares the existing several popular network structures,and combines their advantages,chooses UNet as the basic model to optimize:first change the activation function,expand the neuron perception range,and reduce the neuron deactivation during the back propagation process Phenomenon occurs;secondly,the pooling index method is introduced for up-sampling,reducing the parameters that need to be trained,and improving training efficiency;finally,using the porous space pyramid pooling method to increase the receptive field while concatenating the results with the low-level feature pooling results to extract more Rich semantic features.Comparing the experimental results from both qualitative and quantitative perspectives,it is found that the improved UNet has a better effect on building extraction. |
