| Building extraction from high resolution remote sensing images plays a critical role in natural disaster emergency and managemency,land resource utilization and analysis,and intelligent city construction and planning,etc.With the development of earth observation technology,the automatic and accurate extraction of buildings from high-resolution images is becoming one of the important research topics.Although high-resolution remote sensing images provide a wealth of spectral information,spectral differences between buildings and complex background noise pose significant challenges to building extraction.At present,the building extraction method includes two kinds of methods: traditional methods and deep learning algorithms.Traditional methods need to design features manually,time consuming and laborious,deep learning has become the mainstream in the field of automatic decoding of remote sensing images because of its strong feature representation ability.However,there are still some problems in the deep learning algorithm,such as interior voids and fuzzy edge segmentation.In this paper,the deep learning algorithms are systematically studied to improve the efficiency of information fusion between shallow and deep layers by improving the feature extraction network,and then optimizing the ability of building detail segmentation.The main research contents and achievements of the thesis are as follows:(1)To address the problems of incomplete small-scale building extraction and inaccurate building boundary extraction in building extraction,the residual module of PSPNet feature extraction network is nested with Inception-v3,and then pyramid pooling is performed to improve the network segmentation performance,an we proposed In＿PPM＿Res Net.To verify the effectiveness and generalizability of In＿PPM＿Res Net,experiments are conducted on WHU and AIRS building datasets,respectively,and compared with typical building extraction networks FCN-8s,UNet++,Seg Net,and PSPNet.The experimental results indicate that on the WHU dataset,In＿PPM＿Res Net,Compared to other networks,In＿PPM＿Res Net has the optimal intersection to MIo U of 89.97%.Other evaluation indicators have also been improved to some extent,proving that the efficiency of this model is relatively high.On the AIRS dataset with a resolution of 0.075 m,the intersection to MIo U,Precision,Recall,and F1 score of our model are all the optimal values,reaching 88.51%,86.77%,88.95%,and87.85%,respectively.On the other hand,compared to other networks,the results of building extraction have more accurate boundaries,fewer voids,and certain practicality.(2)To further improve model segmentation accuracy,accelerate model efficiency,and achieve model lightweight,on the basis of U-Net,the spatial attention module of CBAM is embedded in the encoding stage of U-Net,cascaded with the second convolutional layer of each encoding block of U-Net,and a U-Net network combining spatial attention module is proposed.The WHU and AIRS datasets are used to verify the performance of FCN-8s,Seg Net,U-Net,and combined spatial attention module U-Net networks,and the experimental results show that the proposed lightweight model achieves 91.74% and 90.40% scores of accuracy evaluation index MIo U on WHU and AIRS,and compared with FCN-8s,Seg Net,and U-Net,the model in this article,Precision,Recall,and F1 score are also improved to a certain extent.The model in this paper has more complete building extraction results,straighter edges,and low false detection rate and missed detection rate,which prove the effectiveness of the model in this paper.In order to increase the generalizability of this model,the training is assisted by the migration learning method to obtain a convergent model with antiinterference ability.From the results of microscopic quantitative analysis and macroscopic qualitative analysis,it can be seen that on the Inria dataset,compared with the base network U-Net,the evaluation indexes MIo U,Precision,Recall,F1,and m AP of the U-Net network combined with the spatial attention module have been improved,followed by the building segmentation results combined with the spatial attention module U-Net network fit better with the reality and have less hairy edges,and finally the comparison results from the model complexity statistics table show that the model has stronger generalization ability in the acceptable range of the increase in the number of model parameters.(3)To address the problem that the building prediction maps outputted by deep learning-based algorithms cannot be applied directly,a data post-processing study is conducted,the canny edge detection algorithm is used to extract the building outline,and then the raster image vectorization technique with arcpy-based bend＿simplify algorithm is applied to output the building patches vector. |
