Research On Hyperspectral Change Detection Algorithms Based On Tensor Learning

The change detection in multitemporal remote sensing images is one of the valuable tools to monitor the dynamic earth surface.It is widely used in many research activities related to earth observation,such as urban expansion,disaster assessment,resource management and land use.Compared with traditional remote sensing images such as synthetic aperture radar images and multispectral images,hyperspectral images contain both the spatial structure information and a lot of fine spectral information of ground objects.Therefore,the change detection in multitemporal hyperspectral image can detect more subtle changes.However,hyperspectral change detection research also faces many unique challenges.For example,many advanced hyperspectral change detection algorithms often cut and split the spatial and the spectral features.So how to improve the detection performance by retaining the spatial spectral structure correlation of 3D hyperspectral images is one of the design difficulties of hyperspectral change detection techniques.Secondly,due to the high cost of collecting ground reference data,the availability of high-quality labled hyperspectral data sets is typically limited.Thus,it is difficulty for the most advanced supervised machine learning algorithm to be widely used in hyperspectral change detection.In order to solve the above problems,this paper introduces tensor learning into the hyperspectral change detection technique to discriminate between changed and unchanged areas by jointly using spectral and spatial information.Considering the characteristics of hyperspectral image change area,the multi-attribute and multi-scale joint sensing methods is designed to generate pseudo training samples.This method makes pseudo training samples closer to the real labeled samples.The specific work and innovation of this paper are as follows:(1)A hyperspectral change detection algorithm based on tensor learning is proposed.In order to avoid splitting the correlation between the spectral and spatial structure in the change region,this paper introduces tensor learning into the classical ridge regression classifier model.To ensure that the prior information of different modes in the tensor training samples can be effectively learned,it uses CANDECOMP/PARAFAC(CP)decomposition and non negative constraints of different modes so as to construct the tensor ridge regression classifier model.The tensor training samples for this model are get by pseudo training sample generation algorithm.The trained tensor ridge regression classifier model is used to classify the pixels in the hyperspectral difference image,and the hyperspectral change detection results are obtained.Tensor learning can make the model learn the spatial spectral correlation information,thereby enhancing the ability of ridge regression classifier model to distinguish the changed and unchanged pixels.The experimental results show that compared with the classical ridge regression classifier,the detection accuracy of the tensor ridge regression classifier on three hyperspectral datasets is improved by about 0.08,0.06 and 0.01,respectively.(2)A pseudo training sample generation algorithm jointly sensing multi-attribute and multiscale is proposed.It is an expensive and timeconsuming process to collect the ground references for multiple times and for detailed land-cover classes.So it is difficult to obtain the manual labeled training samples.In order to solve this problem,makes comprehensive judgment by leveraging multiple detection methods which are sensitive to the changed objects with different characteristics,and uses the multi-scale the neighborhood-based criterion to screen the alternative positive and negative pseudo training samples respectively,so as to generate multi-attribute and multi-scale pseudo training samples,and then uses the multiattribute and multi-scale pseudo prior information to constrain the classification results of the tensor ridge regression classifier.The multi-attribute and multi-scale joint sensing pseudo training sample generation algorithm improves the reliability of the pseudo training sample data,so that the supervised change detection algorithm can play a better detection ability in the absence of prior information.The experimental results show that the detection accuracy of the hyperspectral change detection algorithm based on tensor learning is improved by0.01 to 0.03,which surpasses the detection performance of the other six contrast algorithms.