Research On Superpixel-based High-resolution SAR Image Interpretation Methods And Applications

With excellent capability of data acquisition and imaging advantages such as penetrating clouds and vegetation,Synthetic aperture radar(SAR)has become one of the main technical means in the field of earth observation today.It also plays an important role in terrain mapping,resource surveying,disaster monitoring,emergency response and national defense construction.With the rapid development of the SAR system,massive SAR data will be obtained in the future,so it is of great practical significance to actively carry out relevant researches on high-resolution SAR image interpretation.As one of the common tools in the field of image processing,superpixel generation technology splits the image to a certain number of semantically meaningful subregions.Comparing to pixels,which is the traditional processing units,superpixels are more conducive to the extraction of local features and the expression of structural information,and can greatly reduce the computational complexity of subsequent processing.In order to improve the performance of high-resolution SAR image interpretation in different applications,this paper combines the superpixel segmentation with traditional pattern recognition and deep learning methods for image interpretation methods and applcations,and the contributions are summarized as follows:(1)To address the problems with multinomial weighted distance applied in superpixel generation of SAR images,this paper proposes an edgedominated superpixel generation method EDLC and also use an adaptive grid to improve the initialization of clusters,thereby improving the performance of superpixel segmentation.Furthermore,based on the densitybased spatial clustering method DBSCAN and the probabilistic patch-based similarity,this paper proposes another superpixel generation method named PPB-DBSCAN,which has more stable performance with fewer iterations.Experimental results on simulated and real images confirm the effectiveness of these two algorithms.(2)To address the problem that the traditional CFAR algorithm cannot accurately model and estimate the backgrounds by using a single model,this paper proposes a method for effectively estimating the background clutter in airport runway areas by means of the generalized gamma mixture distribution,and completes the aircraft target detection in the runway areas.Moreover,the two-parameter CFAR detection algorithm based on superpixel segmentation and the ship target detection network based on Superpixel Net are also proposed for the ship target detection problem in high-resolution SAR images.The experiments with different data prove the effectiveness of these two algorithms for high-resolution SAR image target detection.(3)Considering that most of the traditional change detection techniques of high-resolution SAR image are constructed based on single-pixel information and do not fully utilize the spatial constraint relationship between pixels,this paper proposes a multi-temporal SAR image change detection algorithm based on multi-temporal superpixels and probabilistic patch based similarity after generating superpixels guided by multi-temporal SAR image edge information.Besides,in order to combine the superpixel segmentation with the Auto-encoder,another novel multi-temporal SAR image change detection method is also constructed in this paper.The experimental results show that both methods can effectively identify the change regions in the multitemporal SAR images.(4)For the classification of high-resolution SAR image features,this paper implements superpixel segmention by using EDLC and extracts features of each superpixel such as intensity mean,weighted intensity mean,magnitude mean,weighted magnitude mean,texture mean,texture maximum and weighted texture mean.Then uses superpixels as a node to construct a graph convolutional neural network applicable to the semi-supervised terrain classification of high-resolution SAR images.The results of different data show that the proposed method based on superpixels and GCN perform a better classification than the traditional methods.