Studies On Methods Of SAR Imagery Change Detection Based On Feature Learning

The task of change detection is to confirm the changes and give their quantitative description by analyzing and comparing the images acquired in the same geographical area at two different times.It has been widely used in many applications,like land monitoring,agriculture survey,urban planning,hazard monitoring,map updating,etc.Since Synthetic Aperture Radar(SAR)is not sensitive to cloud cover or sunlight and has all-weather,all-time capabilities,SAR images have become important information resources of change detection.Though the development of space technology makes a large number of SAR images available,besides geometric distortion and radiometric distortion generated by SAR imaging mechanism,speckles produced by the interference of the scatterers' electromagnetic backscattering increase the difficulty of image processing.Speckles mix in weak changes,and weaken the characteristics of the true changes,which affects the performance of the image analysis algorithms directly and causes false alarm and miss alarm heavily.Therefore,by considering the concrete issue of change detection and utilizing the abundant characteristics of the land cover,exploring robust,automatic,accurate and efficient change detection methods is an important research field.After analyzing the key questions of change detection deeply,this thesis makes systematic researches on the two aspects: difference image generation and change detection algorithms.The main achievements are summarized as following.1.Considering the traditional grey difference image is sensitive to speckle noises,a method of constructing grey-texture difference images is proposed.When changes occur between two multi-temporal images on their corresponding land cover,texture also changes.Therefore,texture difference reflects the spatial changes of the images,which is an effective complementary addition to the grey difference image.In the extraction of the texture difference,Robust Principal Component Analysis technique is used to separate irrelevant and noisy elements from Gabor responses.Experimental results show that the traditional pixel-based methods using these grey-texture difference images can find a tradeoff between noises resistance and detail preservation,and give good performance.2.Aiming at grey-texture difference images,a multivariate generalized Gaussian model based graph cuts algorithm is proposed.Though Gaussian is simple,it cannot fit the distribution of the real complex SAR images.Therefore,in the process of construction the data constraint in graph cuts' energy function,generalized Gaussian model is used.And considering the grey-texture difference images,univariate generalized Gaussian model is extended to multivariate one.Such modification makes the algorithm has the ability to approximate a gigantic class of statistical distributions.Change detection experimental results show that the method can guarantee the smoothness of the contours and has more ability to resist noises.3.Inspired by philosophy of manifolds learning,particularly locally linear embedding,a SAR imagery change detection method based on dictionary learning is proposed.This method needs to construct two coupled dictionaries realizing the map from difference image space to the corresponding change-detection map space.The approach can be completed by two steps.Firstly,we approximately represent the input difference image with respect to the difference image dictionary by matching pursuit algorithm.Secondly,the coefficients of this representation are applied with the change-detection map dictionary to generate the output change-detection map.This method is an automatic and off-line learning approach.Therefore,it is suitable for change detection problem such as lacking ground-truth and being difficulty in labeling the data sets.Experimental results show that compared with traditional change-detection techniques,it is an efficient and accurate method.Especially it is good at keeping region consistence,preserving details and resisting noises.4.Aiming at the phenomenon that general deep networks have not utilized the structure similarity of the difference image and its reference,a SAR imagery change detection method is proposed,which realizes the direct map from the difference image input to the changedetection map output by stacked restricted Boltzmann machine.Since hierarchical network enables intermediate features shared and transfer learning possible,the training examples can be obtained by some known images marked in the previous works.Besides,the network constructs the map from image to image by the combination of unsupervised learning and supervised learning.Although the training images are fixed and have no relation to the test images,experimental results show that the performance of this method is excellent,especially in the field of noise robustness.5.In the architecture of deep networks,a SAR imagery change detection method based on patch similarity is proposed.Since the random speckle noises exist in the SAR images,comparing the pixels between the multi-temporal images cannot lead to a correct judgment of their change condition.Therefore,the patches extracted from the multi-temporal images respectively can be described by two convolutional auto-encoders.Then the extracted features are input to the decision network.The output of the decision network is a label related to change or un-change class.At last,the change detection work related to the whole images can be completed based on the patches comparison.Experimental results show that this method gives good performance in accuracy.