Dynamic Graph Neural Network And Meta-Learning For SAR Image Change Detection

Change detection has been widely used in many fields very recently.Its goal is to exploit the change information between a pair of images that observe in the same location at different time points.Synthetic Aperture Radar(SAR)can acquire high-resolution images with sufficient features on the ground surface under any weather and observation distance,which makes SAR images become one of the main data sources in the field of change detection.Nevertheless,there are following challenges in SAR image change detection.1)inadequate extraction of image features by unsupervised methods.2)the poor performance of cross-sensors change detection due to their large differences in distributions.3)imbalance of SAR image data due to a small percentage of change regions.4)high cost of data acquisition in the field of SAR image change detection.Inspired by these challenges,we propose a SAR image change detection method via dynamic graph neural network and meta-learning.The main research contents are divided into the following three aspects:1.To address the problems that SAR images are subject to the speckle noise and the weak ability of unsupervised methods to extract image features,a SAR image change detection based on dynamic graph neural network will be proposed in this paper.A graph construct data is constructed by using two temporal phase images and difference image,and its structure is dynamically updated according to the node features in each training iteration of the training process.This method not only learns the neighborhood information of the image to suppress the coherent speckle noise but also extracts the image texture information and structure information to improve the classification ability of the model.Four sets of SAR image data in real scenes are used to verify the performance of the algorithm.Compared with the detection results of other methods,the proposed method achieves better detection performance on the four sets of datasets.2.To address the problem of data imbalance caused by the low percentage of change regions in SAR images,a SAR image change detection based on semantic data augmentation and meta-learning will be proposed in this paper.The Implicit Semantic Data Augmentation(ISDA)technology can improve the capacity and diversity of training datasets through the process of semantic data augmentation,but the ISDA technique cannot estimate a reasonable intra-class covariance matrix due to the limited number of samples in minority classes,which leads to its poor semantic data augmentation performance.In this paper,we propose to improve the ISDA technique by using MLP as a meta-learner to estimate the intra-class covariance matrix on a small class-balanced dataset.From the change detection results obtained on four sets of real SAR image data,the proposed method can not only achieve better detection performance,but also maintain a high change detection accuracy when the sample imbalance ratio increases further,and the detection performance of the comparison method is significantly reduced.3.To address the problem that the supervised SAR image change detection method requires a large number of label samples,and the label sample acquisition is expensive,a novel SAR image change detection method based on weakly supervised label correction and meta-learning will be proposed in this paper.First,the unsupervised algorithm is used to analyze the difference image to obtain the preliminary prediction results as noisy labels,and then the label correction method is introduced into the change detection task for the first time to avoid the model will be misled by the wrong information in noisy labels during the training process,which leads to the performance degradation of the weakly supervised change detection method.In this paper,we employ a meta-learning method to generate soft labels instead of the original noisy labels during training to replace the original noise labels for training to improve the performance of label correction,thereby further improving the final detection accuracy of the weakly supervised change detection method.From the change detection results obtained on four sets of real SAR image data,the proposed method achieves better detection performance compared with other methods.