Airport Runway Detection With Pol SAR Images Based On Deep Learning

Polarimetric Synthetic Aperture Radar(PolSAR)has the advantage of all-day and all-weather operation,so it can resist the influence of adverse weather to achieve continuous Earth observation.Now the system has been widely used in military and civilian exploration,emergency rescue,target detection and other important fields.Airport runways are the basic infrastructure to support the security of aircrafts,so the automatic airport runway detection technology for Pol SAR images has important research significance.Deep learning algorithms are a branch of machine learning algorithms,and the image processing techniques based on deep learning can extract the deep features of images,which have stronger robustness and better performance in terms of accuracy and efficiency.In this thesis,two complex-valued neural network algorithms are given for airport runway detection by studying the complex data in Pol SAR system’s transceiver channels and the complex-valued neural network.The first algorithm is given a Complex Cross Residual Network(CS-Res Net)based on the combination of Complex Cross Convolutional Neural Network(CS-CNN)and Residual Network(Res Net).Firstly,the given neural network is used to extract the suspected region of interest(ROI)from the Pol SAR images by using the pixel-by-pixel image segmentation method,then morphological processing is performed on the suspected ROI,and finally the geometric features of the airport runway are used to further identify the suspected ROI to extract the real runway area.The second algorithm is given a Complex-Valued Residual Fully Convolutional Neural Network(CV-Res FCN)based on the combination of Complex Residual Network(CV-Res Net)and Fully Convolutional Network(FCN)to extract the airport runway more efficient and accurate.Firstly,the given algorithm is performed by down-sampling and up-sampling operations to give independent feature vector for each pixel in the input image,then the Softmax classifier in the network is used to probabilistically predict the class of each pixel’s feature vector to extract the suspected ROI,finally the prominent parallel line features and topological features of the runway is used to identify the ROI to obtain the real airport runway region.In order to verify the performance of the two Pol SAR image airport runway detection algorithms given in this thesis,several sets of comparison experiments are executed on measured Pol SAR data in this thesis.The result images show that compared with the comparison algorithms,both methods given in this thesis not only can completely detect the airport runway region from the image but also maintain better details of the runway,and have better robustness.