| Remote sensing image change detection is a significant technology to identify the changed regions through multitemporal remote sensing images acquired on the same geographical area at different times.In recent years,heterogeneous remote sensing image change detection has received widespread attention due to its advantages,such as no need for labels,no restrictions on the types of input remote sensing images,and easiness of implementation and usage.At the same time,deep neural networks,which achieved large improvements in the field of computer vision,have also been applied in remote sensing image change detection problems due to their properties of deep abstraction and automatic representation learning.Focusing on the difficulties in heterogeneous image change detection problem,this thesis proposes three unsupervised heterogeneous image change detection methods based on existing change detection methods and deep neural network technology.The main researches in this thesis are listed as follows:(1)An unsupervised heterogeneous image change detection method based on coupling convolutional autoencoder and mapping transformation is proposed.Firstly,each pixel with its neighborhood will be transformed into a feature vector through the coupling convolutional autoencoder to eliminate the redundancies and obtain more consistent feature representations.Secondly,the method builds a mapping neural network to explore the inner relationships between the feature vector pairs.Thirdly,the learned mapping function can bridge the different feature representations,and the feature vector pairs will be mapped into a consistent feature space.In the feature space,the feature vector pairs can be compared and the difference map can be calculated.Finally,a clustering algorithm is used to analyze the difference map and generate the change map.The coupling convolutional autoencoder uses the convolutional neural networks to extract features in input images such as color,shape,texture,and spatial relationships.It eliminates the redundancies in the raw images and weakens the effect of noises.The experiments show the effectiveness of the proposed method.(2)An unsupervised heterogeneous image change detection method based on commonality autoencoder with negative samples elimination is proposed.Firstly,each pixel with its neighborhood will be transformed into a feature vector through the coupling convolutional autoencoder.Secondly,the method designs a commonality autoencoder to learn the mapping function between different feature representations and eliminate the negative samples from regions with a high probability of change.Since the samples from changed regions do not participate in the training process,the learned mapping function will reduce the difference of unchanged regions and highlight the difference of changed regions.The experiments show that the proposed method effectively improves the quality of the difference maps and the method can generate more accurate change detection results.(3)An unsupervised heterogeneous image change detection method based on dual commonality autoencoder is proposed.Firstly,each pixel with its neighborhood will be transformed into a feature vector through the coupling convolutional autoencoder.After that,the method designs a dual commonality autoencoder to measure the difference between the feature vector pairs directly to generate a difference map.The dual commonality autoencoder uses the bi-directional consistency of change detection to construct a more robust difference measure standard.It integrates the feature difference measured by the two commonality autoencoders to calculate the total feature difference.The method improves the robustness and the quality of the difference map.The experiments show the effectiveness and applicability of the proposed method. |
