Polarization SAR Building Area Extraction Method Based On Deep Convolutional Neural Network

Identification and extraction of built-up areas,as main content of surface features category and important mapping elements in topographic map,are becoming more and more important along with rapid development of cities and construction of digital cities.Although favorable research achievements have been made with regard to extraction of built-up areas in optical images,optical sensor imaging depends on light condition and it couldn't realize sustained and effective monitoring in misty,rainy,cloudy,snowy weather and nightly environment.With advantages of both microwave band initiative remote sensing and multi-polarized synthetic obtaining of surface feature information,Polarimetric Synthetic Aperture Radar(PolSAR)is capable of expressing and monitoring surface features more exactly than traditional optical images.As a result of the special imaging method of PolSAR images,there are plenty of complicated scattering characters in the built-up areas in PolSAR images,which greatly increases the difficulty of identifying the built-up areas in PolSAR images.At the meantime,people's demand of rapid automatic extracting built-up areas from mass data couldn't be satisfied by the traditional means of interpretation as the SAR system is continuously improving and developing.As the deep learning method has been rapidly developing in the image processing field and revealed excellent result over the past few years,more and more researchers have focused on how to better integrate deep learning with SAR images.Therein,the semantic segmentation approach within deep learning can be utilized for carrying out per-pixel classification of images so as to achieve favorable image segmenting results.The semantic segmentation network based PolSAR image built-up area extracting method can not only utilize the polarized scattering character of PolSAR data but also extract the high-rise features of the images so as to get the information of multidimensional surface features engaged in the process of extracting the built-up areas,which makes for improving the feature utilization rate of the PolSAR data.On such basis,this paper carried out exploration as the followings:(1)Proposed a method that integrates full convolutional neural network models(FCN)and conditional random fields(CRF)for extracting built-up areas in PolSAR images.This method targets the three components produced by decomposition of Freeman respectively to the three input channels of the network and makes full use of the low-rise features of PolSAR images and the network characters of semantic segmentation of FCN to realize end to end classification.Meanwhile,it utilizes CRF to perform post-optimization treatment for better connecting the contextual information,thus the extracting results could be more precise.Experiments show that the integration of FCN and CRF can not only improve the extracting precision of the built-up areas,but also greatly reduce blending of the built-up areas and the other surface features.(2)Proposed applying the new network LinkNet in the semantic segmentation network to PolSAR images.As a codec structural network consisting of multiple coders and decoders,this network can also relate the preamble input information to apply this network in extraction of built-up areas in PolSAR images.It has verified the effectiveness of deep learning network model of various structures on extraction of built-up areas in PolSAR images,and it has also explored the matters needing attention in making sample set of PolSAR images with regard to different network structures.Experiments showed that more than precisely extracting built-up areas in small images,the LinkNet can also achieve favorable results in extracting built-up areas in large images.(3)Carried out detailed comparison between the two methods proposed in this paper and the traditional PolSAR image processing methods H/a-Wishart and SVM.Analysis showed that in spite of the two methods proposed in this paper had omissive conditions within small ranges,their overall precision were higher than that of the two classic methods,especially had favorable built-up area extracting completeness and outline extracting precision,and they could realize automatic extraction at a certain degree.