Research On Building Extraction From Remote Sensing Images Based On Fully Convolutional Neural Network

Buildings are one of the important artificial features in the city.The extraction of building information based on high spatial resolution remote sensing images plays an important role in the intelligent census of urban buildings,real estate registration,and urban construction.Due to the complexity of the building itself,the traditional extraction method cannot efficiently and accurately complete the automatic extraction of the building.Therefore,how to accurately and quickly realize the automatic extraction of the building in the high-resolution image has become an urgent problem to be solved.Among the traditional extraction methods,although the visual interpretation method has high extraction accuracy,the extraction process is time-consuming and labor-intensive,and is not suitable for large-scale promotion.Compared with visual interpretation methods,traditional machine learning and convolutional neural network algorithms have significantly improved extraction speed,but the extraction accuracy is relatively low.In view of the shortcomings of the above research,this article first conducts experiments on the commonly used full convolutional neural network algorithm,determines U-Net as the basic network model of this article,and then introduces different forms of attention mechanisms on the basis of U-Net to construct SE-Unet,CBAM-Unet and SK-Unet three building extraction network models;finally based on the SK-Unet network model,using data-dependent upsampling to construct the SD-Unet network model used for building automation extraction in this article.The main research content and results of the paper are as follows:(1)Using the data set Ⅱ in the satellite building database of the WHU data set as the original data,and using data enhancement methods to establish a building data set suitable for this study.(2)In order to verify the extraction performance of the commonly used fully convolutional neural networks U-Net,SegNet,and LinkNet,experiments are carried out on the data set produced in this article,and the basic network model of this article is determined.The experimental results show that:U-Net network model with 74.84%intersection ratio,83.72%recall rate,84.95%F1 score and 98.85%overall accuracy become the optimal model among the fully convolutional neural network models commonly used in this article,so This paper chooses U-Net as the basic network model,and improves and optimizes it on the basis of U-Net model to improve its ability to extract buildings.(3)Aiming at the commonly used full convolutional neural network due to the fixed size of the neuron acceptance domain and the limitation of short-range context,three different forms of attention mechanisms are introduced to construct SE-Unet,CBAM-Unet and SK-Unet building extraction networks,And experiment on the data set.The experimental results show that:compared with the U-Net network model,the SE-Unet,CBAM-Unet and SK-Unet network models have significantly improved the accuracy of building extraction.Among them,the SK-Unet model that introduces the convolution kernel attention mechanism is in The cross-to-bin ratio,recall rate,F1 score,and overall accuracy are the best,respectively:75.90%,87.33%,85.55%,98.81%.Through comparative analysis of the PR curve and ROC curve of the SE-Unet,CBAM-Unet and SK-Unet models,it is confirmed that the SK-Unet network has superior extraction performance.(4)In view of the limited ability of traditional bilinear upsampling to recover and predict,data-dependent upsampling is introduced on the basis of SK-Unet,an SD-Unet model is constructed and experiments are performed on the data set.The experimental results show that the SD-Unet model achieves better extraction accuracy on the experimental data set,and the cross-to-bin ratio,F1 score,accuracy,and overall accuracy are 76.25%,85.85%,86.13%,and 98.86%,respectively.Draw the PR curve and ROC curve of the model by using the predictive probability graph of the SD-Unet model and compare and analyze the PR curve and ROC curve of other models.It is found that SD-Unet is the model with the best performance among the network models mentioned in this article.It is proved that the data depends on upsampling and has a strong ability to predict recovery.(5)Aiming at the rough edges of buildings extracted by the full convolutional neural network,the curve compression and minimum vertex compression algorithms are used to optimize the edges of the buildings to make the overall outline of the building more regular,which is more conducive to the update and use of the electronic map of the building.