Automatic Change Detection Method Of SAR Images By Integrating The Intensity And Coherence Information

There is great significance in managing natural resources,protecting the ecological environment,and researching social development to understand the precise and real-time biophysical changes that occurred on the Earth’s surface caused by natural processes and human activities.As an active microwave imaging system,Synthetic Aperture Radar(SAR),with the capabilities of all-time and all-weather observation,wide detection range,rich image information,and so on,has become a vital technological means for remote sensing change detection.However,change detection technology still faces many problems and challenges,such as low registration precision for multisource images,poor applicability of detection methods,and low degree of automation.Therefore,the dissertation,taking the spaceborne SAR images as the research object,proposes a high-precision registration method for multi-source SAR images,and based on this and making full use of statistical characteristics of images,intensively studies high-precision and automatic change detection methods,thus providing efficient solutions for SAR image interpretation and application.The main research work and contributions are as follows:(1)A new registration method assisted by Range-Doppler(RD)model for multisource SAR images is proposed by fully utilizing imaging geometric information and geometric distribution properties of pixels of SAR images.The method is improved in feature extraction,feature matching,and geometric transformation.Firstly,a spatial correlation strategy based on stationary wavelet transform(SWT)is introduced,improving the stability of feature points.Secondly,RD model is introduced to assist feature matching,which turns the matching mode from global to local and improves the matching accuracy and efficiency.Finally,an adaptive geometric transformation model is established by utilizing RD model to convert coordinates from slant range to ground range,effectively weakening the geometric distortion differences caused by topographic relief between images and improving the geometric transformation accuracy.The experimental results on SAR images in different imaging configurations show that the proposed method can obtain sufficient matching features with good spatial distribution,especially on SAR images with significant geometric differences,indicating the applicability to multisource images.And it has higher execution efficiency compared with the same type of methods.(2)An adaptive and adjustable method of intensity-based change detection for SAR images is proposed.On the basis of the theory that SAR complex images obey the circular Gaussian distribution with zero mean value,the method derives an intensity change statistic based on the statistical characteristics of SAR images.It is theoretically proved that the statistic has the adaptive advantage,automatically determining the stretch threshold for each pixel.Meanwhile,superpixels segmentation is introduced into the statistic estimation process,which makes the method adjustable with the ability to detect different degrees of change.The experimental results on SAR images with different statistical distribution characteristics show that the suggested statistic can generate the difference image of quite obvious contrast,effectively improving the distinction between changed and unchanged areas.Moreover,the proposed method achieves better change detection results and has better applicability by reaching a better compromise between false alarms and missing alarms and reducing the number of error detection compared with other classical thresholding methods.(3)A coherent change detection(CCD)method of SAR images based on the weighted coherence change statistic is proposed.Based on the statistical characteristics of SAR images,a weighted coherence change statistic is derived by introducing the ratio change statistic representing intensity differences as weights,which avoids the limitation of equal variance assumption and improves the contrast between changed and unchanged areas,making it more suitable for SAR change detection tasks.The introduction of weights takes into account both coherence and intensity change in the detection process and improves the accuracy of change detection.Besides,compared with the state-of-the-art coherence statistics,the weighted one provides lower coherence values in changed areas and obtains the higher-contrast difference image in changed and unchanged areas with much clearer boundaries.(4)Intensity-based change detection methods have advantages in detecting large-scale changes,while CCD methods tend to identify small-scale changes.To simultaneously detect two scale changes,a change detection method is put forward based on the statistical characteristics of SAR images and improved Res Net(I-Res Net),which realizes the automation from preclassification and sample selection to model training and prediction,forming a complete change detection processing flow.The proposed method takes the fusion result of change maps obtained by the suggested intensity-based change detection method and CCD method as the preclassification image.On this basis,the sample numbers are allocated proportionally by dividing changes into positive change and negative change,improving the generalization ability of the features extracted by the network in changed areas.An I-Res Net model is designed by combining the advantages of Res Net in easy training and avoiding information loss and wavelet transformation in noise reduction.The experimental results show that the I-Res Net method obtains good detection results with small sample training sets,improving the training efficiency,and can detect two scale changes by identifying large-scale changes while keeping more subtle changes.Finally,a comparison is performed between the three proposed methods,pointing out their advantages and disadvantages and application conditions.There are 86 figures,12 tables,and 201 references in the dissertation.