Study On Deep Learning-Based SAR Images Change Detection

Possessing ability of working at all-day and all-weather situation makes synthetic aperture radar(SAR)one of the trends in remote sensing area.And change detection of SAR images is one prevalent application of SAR,whose purpose is to gain change information through analyzing SAR images taken from one area at different times.Change detection has been utilized to disaster assessment,urban planning,crop monitoring and military investigation.Based on the existing SAR image change detection methods,combined with new theories such as deep learning and superpixel segmentation,this thesis proposes a new SAR image change detection method.The main work of this thesis is as follows:(1)Aiming at the uncertainty of the existing deep learning change detection methods,which stems from the generation of training patches by pixel-wise,this thesis proposes an algorithm of SAR image change detection based on wavelet convolution neural network(CWNN).Firstly,the original SAR images are segmented by superpixel segmentation,and the superpixels are used as the objects to classify and extracted as samples for network training.Compared with the traditional methods of taking pixel as the target to extract samples,superpixels segmentation can take homogeneous region as the units of classification by adding neighborhood information,and select more recognizable training samples for the network in an object-oriented way.Then the extracted samples are sent to CWNN for training.Compared with convolution neural network(CNN),CWNN's pooling layers are wavelet pooling layers,which can pool some high-frequency information.It has a better effect on suppressing speckle generated by SAR imaging mechanism and reducing its impact on change detection results.Finally,the trained network is used to classify the superpixels which are not selected as training samples to obtain the final change map.The experiments on three sets of real SAR images show that the proposed method has 14.15% higher accuracy and 26.42% higher kappa coefficient(KC)than the other two algorithms at the most,which proves that the proposed algorithm has better performance.(2)To repress the issue rooting in strong speckle noise of SAR images that detection results reveal high false alarm ratio,this thesis proposes an algorithm of two-phase SAR image change detection based on PCANet.In the first phase,the original SAR images are segmented by superpixel segmentation,and the extracted samples are sent to pcanet1 for training.PCANet uses PCA filter banks as hidden layers instead of convolution layers in CNN,so in the process of network training,there is no need for parameters regularization and optimization solver,which improves the efficiency and accuracy of the network.The trained PCANet1 is used to classify SAR images into two categories,one is unchanged class,the other is changed class.The changed class includes changes caused by noise and changes caused by real objects.In the second phase,only the pixels of changed class are classified.Low rank sparse decomposition is used to reduce the influence of speckle noise on superpixels,and then samples are sent to PCANet2 for training.Finally,PCANet2 after training is used to detect the image and gain the final change map.The experiments on three sets of real SAR images show that the proposed method has 14.15% higher accuracy and26.42% higher kappa coefficient(KC)than the other two algorithms at the most,which proves that the proposed algorithm has better performance.