Research On Hyperspectral Change Detection Based On Unsupervised Network

The change detection technology of remote sensing images analyzes the changes of the same scene at different times to realize the interpretation of remote sensing scenes and obtain the information about the large-scale changes on the earth’s surface.It has important applications in the field of remote sensing image processing.Compared with panchromatic and multi-spectral images,hyperspectral images can characterize the spatial information of the features and higher resolution and finer spectral information.So hyperspectral change detection plays an important role in the field of resource and environmental monitoring,natural disaster assessment,military reconnaissance and so on.However,there are still many challenges to achieve effective hyperspectral change detection.For example,how to mine the discriminative deep features while reducing the dimension of the hyperspectral data to solve the problem of the unsatisfactory accuracy caused by noises or the other interference information.How to design a deep network architecture with constrained spectral distance and strong data representation capabilities to simulate a large range of complex hyperspectral remote sensing scenes effectively.In addition,most of the current hyperspectral change detection algorithms based on deep learning are supervised networks,but the selection and evaluation of the hyperspectral training samples requires more manpower,resources and time.So the samples are very scarce.In view of the above problems,this paper analyzes the data characteristics of the multitemporal hyperspectral images.Based on the current hyperspectral change detection technology,we explore a network architecture which can effectively implement hyperspectral change detection.We use the real multitemporal hyperspectral remote sensing data to evaluate the model in this paper.The research contents and innovations of this paper are as follows:Traditional transform detection methods cannot characterize the deep-level structural relationships of the multitemporal hyperspectral images.At the same time,the problems of insufficient training samples and difficult manual labeling are considered in hyperspectral remote sensing images.We propose a deep-level feature extraction method for multi-temporal hyperspectral images based on unsupervised auto-encoder networks,which not only reduces the dimension of the hyperspectral data but also mines the discriminative features of changing regions.Unlike other remote sensing images,such as panchromatic and multispectral images,the spectral information of hyperspectral images can essentially distinguish targets.It is the key attribute.In order to further improve the accuracy of change detection and mine the essential spectral characteristics of the hyperspectral images,this paper introduces an unsupervised network with the constraint of spectral distance through adversarial learning.By comparing the spectral angular distance between the original data and the non-linearly mapped data,it proves that the network can effectively characterize high-dimensional spectral data in deep latent space.Comparing with the supervised networks,experimental results show that the detection accuracy is improved by about 0.01 to 0.02 based on such features.Based on the above-mentioned deep-level spectral features in the depth-latent space,this paper also designs a spatial feature extractor based on a multi-scale attribute filter to achieve effective cascading of spatial-spectral features to ensure multi-scale and multi-shape change region discrimination accuracy.The basic idea is to build a multi-scale morphological attribute open-operation and closed-operation spatial feature extractor based on deep-level spectral features extracted from an unsupervised adversarial network.In this way,it is not only convenient to identify the change regions with essential attribute differences in the spectral dimension,but also to distinguish the change regions of different scales and shapes in the spatial domain.