Deep Learning Based Hyperspectral Image Change Detection

Various human activities and changes in the nature are slowly changing the landscape every day.The rapid development of modern high technology,the population growth and the process of urbanization are undoubtedly accelerating this change.Therefore,how to make use of remote sensing image data to timely and effectively monitor surface changes and achieve dynamic monitoring of various fields in life,is of great significance.This is also an important research content in modern remote sensing field.When using multitemporal remote sensing images for change detection research,hyperspectral images have unique advantages.As an image cube that combines two-dimensional image space and one-dimensional spectral information,each pixel in the hyperspectral image is recorded as an approximately continuous spectral characteristic curve.The curve can show different radiation intensities of different ground features.The original hyperspectral image contains a lot of noise and is highdimensional,which makes it difficult for traditional image processing methods to extract features of ground features.Considering that deep learning has strong nonlinear fitting and learning capabilities,it can effectively process high-dimensional data,and then efficiently handle hyperspectral image change detection tasks.Therefore,this paper uses the powerful feature expression capabilities of deep learning to design change detection algorithms for hyperspectral images.The main contributions are as follows:1.Aiming at the problems that there are some mixed pixels and few change detection data sets in hyperspectral images,this paper proposes an end-to-end hyperspectral image change detection framework based on deep affinity matrices.First,the hyperspectral images are linearly and non-linearly mixed to obtain different types of abundance maps.Then the spectral information and the sub-pixel level information of the abundance map is fully fused to construct mixed-affinity matrices.The mixed-affinity matrix can naturally map the differences between two one-dimensional spectral vectors and two one-dimensional abundance vectors into a twodimensional matrix,mining the differences between two different spectral bands and the two abundance features.The mixed-affinity matrix maximizes the utilization of multi-source information and provides cross-spectrum feature information.Due to the small number of hyperspectral image change detection datasets,an efficient lightweight convolutional neural network was designed to prevent over-fitting of the deep network.The constructed mixedaffinity matrices are input into the deep network for feature learning.The last layer outputs the change detection results.The network framework can use both spatial and spectral information,effectively improving the performance of change detection.2.Aiming at the problem of insufficient data sets due to the difficulty in labeling the hyperspectral image in change detection task,this paper proposes a hyperspectral image change detection framework based on unsupervised noise variational modeling.This framework uses some common unsupervised change detection algorithms to generate change detection results with noise,and supposes the noise obeys Gaussian distribution.The change detection map without noise is defined as the hidden variable of the model.Through the noise reduction model based on variational inference,the noise and real state are modeled.Finally,the real change detection results without noise are output by training the fully convolutional network(FCN).Specifically,the proposed network framework consists of three modules: FCN-based feature learning module,two-stream feature fusion module and unsupervised noise modeling module.These three modules cooperate with each other and optimize in an end-to-end way to jointly improve the performance of change detection method.