Unsupervised Deep Representation Learning For Remote Sensing Image Change Detection

Remote sensing image change detection is a key technology in the earth observation applications such as environmental monitoring,agricultural survey,urban research,forest resource monitoring,and disaster assessment.It is of great significance to study dynamic relationships between human activities and the earth’s natural environment.In recent decades,with the continuous development of remote sensing imaging technology,the spatial and temporal spectral resolution of remote sensing images has been continuously improved,showing the characteristics of multi-source,heterogeneous,and massive.In addition,compared with other remote sensing image interpretation technologies,change detection processes and analyzes remote sensing data obtained from the same geographical location at different times,which faces more data volume,stronger heterogeneity,more complex distribution of ground objects,and high-quality annotation data is more difficult to obtain,etc.In order to overcome these difficulties and meet the growing application needs for speed,accuracy,robustness and stability,this thesis considers the characteristics of remote sensing data,takes deep learning as the main method,and focuses on the exploration of unsupervised techniques,deeply studied the change detection problem in the following aspects.1.A novel CNN-based model is proposed for change detection in synthetic aperture radar images.The change detection task uses image pairs as input,which can be regarded as an image pair similarity measurement problem.Based on this,Siamese convolutional network and its variants Pseudo-siamese,2-Channel network are explored to solve SAR image change detection problem.In order to better share information between image descriptors and learn the semantic difference information between changed and unchanged pixels more effectively,we design Siamese-sample convolutional network.The network extracts the descriptors from image pairs as undifferentiated samples,and combines the extracted descriptor pairs to complete the classification using a Softmax classifier.In order to avoid the use of manually labeled data,the network is trained by improving the FCM-based joint classifier to obtain pre-classification results,which then are used as fake labels.Experimental results show that the method can effectively detect changes,and Siamese-sample network achieves a better balance between accuracy and speed among its counterparts.2.A temporal prediction-based self-supervised method is proposed for remote sensing change detection.Self-supervised representation learning usually extracts features beneficial to its target task by mining its own data structure information as free supervised information.Inspired by generative adversarial networks,we mine the temporal information of change detection data as supervisory information.Analogous to the discriminator in generative adversarial network,we train a network to distinguish input samples between time 1or time 2.Through this self-supervision mechanism,the discriminator network can hardly distinguish the samples from common distribution,so it can reduce the distance of the common distribution;for non-common distribution,it has different characteristics,the network can distinguish it,so the distance between non-common distribution is enlarged.Such the learning mechanism can obtain more consistent representations of two images and suppress noise effectively.Finally,the detection result is generated by directly comparing the obtained representations.Experiments verify the effectiveness of the method on multiple real remote sensing datasets.3.A multi-scale self-attention deep clustering is proposed for SAR image change detection.Due to its powerful feature extraction ability,convolutional neural network has achieved great successes in many fields,such as semantic segmentation,object detection,image classification,remote sensing interpretation and so on.However,convolutional neural network requires a lot of labeled data to train.The clustering algorithm utilizes the similarity measurement between data to complete image classification without supervision.However,raw data has not always been cluster-friendly,often resulting in less robust clustering results.Therefore,unsupervised clustering is combined with the convolutional neural network to jointly learn the clustering arrangement of the resulting features and the parameters of the neural network.The method first uses K-means++ algorithm to cluster untrained deep features,and then uses the obtained clustering arrangement as supervisory information to update the network weights.This process makes full use of the complementary property of deep networks and unsupervised clustering,which can promote each other to obtain better solutions in an iterative process.In order to suppress noise interference in the joint optimization process,octave convolution and self-attention mechanisms are introduced to increase the robustness of the network to noise.In addition,multi-scale input information is fused by the designed multi-scale fusion module to extract more refined contextual semantic information for each pixel.Experiments on real SAR datasets verify that the algorithm can effectively extract cluster-friendly deep features and produce more accurate clustering results.4.A deep shearlet convolutional network is proposed for change detection in SAR images.Convolutional neural network can extract shift-invariant features,but it is sensitive to speckle noise in SAR images.In order to alleviate this problem,we introduce shearlet transform into convolutional neural network.In order to retain more texture details while suppressing noise,a shearlet denosing layer is proposed,which combines shearlet transform and threshold denoising algorithm.In this layer,convolutional feature maps are decomposed into high frequency subband and low frequency subband information with shearlet transform.Since the low-frequency subbands contain more energy information,they are delivered to subsequent layers to capture robust high-level features.The high frequency sub-band coefficients usually contain noise,so the hard threshold shrinking algorithm is used to discard the noise of small coefficients,so as to achieve the purpose of noise reduction.Based on this module,a deep shearlet network is designed suitably for SAR image change detection.In order to train the network in an unsupervised manner,the imprecise pre-classification is novelly regarded as noisy labels,and the noise robustness loss is introduced to train the network,which effectively avoids the interference of noisy labels on network optimization.Experiments verify the effectiveness of this algorithm on multiple remote sensing datasets.5.As for the problem of change detection in large-scale high-resolution SAR images,a Transformer-based contrastive representation learning framework is proposed.Compared with low-and medium-resolution SAR remote sensing images,high-resolution images have more complex ground objects and suffer from more serious speckle noise.To this end,we construct a contrastive representation learning framework and adopt the Transformer structure as the backbone network to learn discriminative feature representations.This unsupervised framework learns feature representations from unlabeled data,and an efficient window-based Transformer is used to encode raw image data into hierarchical high-level semantic features.To model rich local-global features,a convolution-enhanced module is designed to establish cross-window information interaction,which effectively enhances the representation ability of window-based Transformer while maintaining computational efficiency,so that more abundant spatial information can be modeled.To improve the adaptability of the algorithm to different data,the features of different layers are used to analyze the difference level,and the decision-level fusion is designed to obtain the fusion change detection result.In addition,to reduce the computational and memory requirements caused by large scale,a sparse sampling strategy is designed,which effectively improves the training and inference speed of the algorithm on high-resolution large-scale SAR images while maintaining high accuracy.