Research On Change Detection Method For SAR Images With Small Samples

As an active microwave imaging sensor,synthetic aperture radar(SAR)can observe the ground scene in a wide range all day and night.It plays an irreplaceable role in urban planning,disaster monitoring,and land use.With more and more spaceborne SAR and airborne SAR sensors being put into use,the amount of available SAR image data is also increasing sharply.In particular,since the orbit of spaceborne SAR sensor is fixed and images can be obtained in a certain period,the SAR image data of the same geographical scene obtained at different time points is also increasing,thus research on the automatic change detection algorithm attracted great attentions.Since the change detection algorithm is often applied to complex scenes,such as flood and landslide areas caused by rainstorm,secondary disaster caused by earthquake,the changes caused by human activities,and a large amount of speckle noise will be generated in the acquisition process of SAR image,these factors together lead to the difficulty of change detection task.How to overcome the interference caused by the complex background and the speckle noise,so as to detect the region of interest accurately is a challenge in SAR image understanding and interpretation.This dissertation takes multi-polarization airborne SAR image and single-polarization spaceborne SAR image as the experimental data.Aiming at the small sample problem in SAR images(i.e.,the difficulty of obtaining ground truth labels in change detection),as well as the inherent problems of speckle noise and complex background interference in the image,by improving of the description method of change information,the feature learning and classifier design,this dissertation deeply studies the SAR image change detection technology based on the deep learning method,and proposes a variety of change detection algorithms.The core research work of this dissertation can be summarized into the following four parts:Considering that the complex scattering mechanism and inherent speckle noise in polarimetric SAR image seriously hinder the change detection task,and the difficulty of accurate detection of small-size changed vehicle targets,we studied the unsupervised change detection algorithm of polarimetric SAR image,and proposed a polarimetric SAR image change detection method based on convolutional neural network under the framework of preclassification.In the preclassification stage,through making full use of the spatial information and scattering information of polarimetric SAR images,a difference map calculation method based on joint superpixel segmentation and region growing is proposed,followed by which the generated difference map is clustered to realize the initial change detection of bitemporal polarimetric SAR images,providing reliable pseudolabels for subsequent model training.By using the pseudolabels and the scattering information of polarimetric SAR image,the designed convolutional neural network can be trained in a supervised manner.Finally,the trained convolution neural network is utilized to predict the label of each pixel in the entire image to obtain more accurate classification results.Due to the combination of spatial information and the scattering characteristics of polarimetric SAR data,this algorithm can effectively suppress the interference from speckle noise and the targets such as buildings and trees in the scene,thereby improving the accuracy of change detection.The experimental results show that this algorithm is more robust than the traditional SAR image change detection algorithm.Aiming at solving the problem of lacking prior label information in bitemporal SAR images,a semi-supervised SAR image change detection method with adversarial learningconstrained co-training is proposed.Firstly,a difference image reflecting the change degree of each pixel position is calculated by the log-ratio operator,and then the shallow input feature is constructed by using the original bitemporal SAR image and the log-ratio difference image.The constructed feature can be used as the input of the network for the extraction of high-level features.Secondly,by constructing two generative adversarial networks and combining the idea of co-training,the proposed model is restrained to extract meaningful category information from a large number of unlabeled data as pseudolabels,so as to expand the training set of the model and finally improve the change detection performance.Through the proposed semi-supervised learning method,in the context of small labeled training samples,the model can still excavate the information in unlabeled data and employ them to ensure the change detection performance.Aiming at the problem that the conventional measurement fashion of change information lacks spatial context information and is seriously affected by speckle noise,a label consistent self-ensembling change detection method is proposed to solve the small labeled sample problem.Under the framework of deep semi-supervised learning,we designed two shallow features which are suitable for SAR image change detection,namely dual features.The dual feature includes two parts: pixel-wise feature and context-wise feature.The pixel-wise feature is generated by using the data of the log-ratio difference image,which can effectively describe the edge and details in the image.The context-wise feature is generated by a difference image combined with spatial context information,which can effectively describe the information of significant change area in the image and suppress the speckle noise and difference value fluctuation in the background area.Considering the complementarity of the dual features,a two-stream label consistent self-ensembling network is designed,and highlevel features can be extracted from the two types of features to improve the representation ability of the network.Then,by fusing the prediction probability of two-stream networks,we can obtain a label prediction that integrates detail information and context information.At the same time,by integrating the network prediction at different training epochs,the prediction of samples can be further improved.More importantly,by fusing the predictions of dual feature information and the predictions of different training epochs,the network can effectively transform the unlabeled samples into pseudolabeled samples,thus expanding the training data of the deep network and improving the change detection performance.Aiming at the problem that deep neural network depends heavily on prior category labels and hybrid variability of changed regions,a hybrid variability-inspired SAR image changed detection model is proposed.Changed regions always tend to have different sizes,irregular shape and texture,that is the hybrid variability.We proposed a new processing unit for the input of deep neural network under the multiresolution analysis framework,which extends the traditional single-scale and single-channel image patch into multi-scale and multichannel image patch,thereby enhancing the feature description of each pixel position in the image.The constructed processing unit can better describe the changed region and overcome the hybrid variability problem.Then,a hybrid variability aware network is proposed under the framework of self-supervised learning.By automatically mining the underlying data structure within abundant unlabeled samples,the proposed network can effectively extract the useful knowledge for change detection task,and realize an end-to-end training.Through the effective learning of multi-scale and multi-channel image patch input,the network can extract and integrate the local spatial structural features and multiscale-multiresolution features,thus effectively enhancing the separability and expression ability of high-level features.More importantly,a self-supervision layer is embedded behind the output layer of the network,which can classify high-level features and construct supervision signal,thereby promoting the network learning in a self-supervised manner.