Research On Sub-pixel Mapping Of Remote Sensing Imagery Via Bayesian Framework

Due to the limitation of hardware and the influence of environment,there are a large number of mixed pixels in remote sensing images,which restricts the application of remote sensing images in Earth observation missions.The spectral unmixing only obtains the abundance of the mixed pixel,but does not obtain the specific spatial distribution of each class in the mixed pixel.As a follow-up processing step of spectral unmixing,sub-pixel mapping is a technique that can obtain higher spatial resolution classification maps.In this paper,the present problems of sub-pixel mapping are studied by using Markov random field and Conditional random field under Bayesian framework:(1)In order to adjust the abundance and endmember adaptively,the dependence on algorithm initialization is reduced.In this paper,an unsupervised Bayesian sub-pixel mapping of hyperspectral imagery based on Band-Weighted Discrete Spectral Mixture Model and Markov Random Field is proposed.The Band-Weighted Discrete Spectral Mixture Model is used to adapt to the noise heterogeneity in hyperspectral imagery and the hidden label field of subpixels,and it is integrated with Markov random field into Bayesian framework.At the same time,the Bayesian framework is optimized by the designed expectation-maximization algorithm,and the endmember and label field of subpixels are estimated by iterative method,which can reduce the initialization dependence of the algorithm.The effectiveness of the proposed method is proved by simulated,real and synthetic data experiments.(2)In order to consider the problem of endmember variability in sub-pixel mapping,a Bayesian sub-pixel mapping of hyperspectral imagery via Discrete Endmember Variability Mixture Model and Markov Random Field is proposed.The Discrete Endmember Variability Mixture Model uses the endmember-abundance model of hyperspectral imagery to adapt to the endmember variability of hyperspectral imagery and the discrete label field of subpixels,and integrates it with the discrete label field modeled by Markov random field into Bayesian framework.Thus,spatial and endmember variability information can be effectively used in subpixel mapping.Through simulated and synthetic data experiments,the proposed method has achieved ideal results.(3)In order to reduce the influence of abundance error on sub-pixel mapping while ensuring operation efficiency,a Bayesian sub-pixel mapping method based on Spatial Adaptive Attraction Model and Conditional Random Field is proposed.The Spatial Adaptive Attraction Model changes the display form of abundance constraint into an implicit form and models it with the spatial smoothing prior in the Conditional random field,so that,the sub-pixel mapping can not only maintain the operation efficiency,but also deal with the noise artifacts,and weaken the influence of abundance error on the sub-pixel mapping results.The effectiveness of the proposed method is illustrated by synthetic and real data experiments.(4)In order to reduce the influence of abundance error on sub-pixel mapping while considering large-scale spatial correlation information,a Bayesian sub-pixel location method based on Deep Full Convolution Neural Network and Conditional Random Field is proposed.The U-Net deep full convolution neural network with jump connection structure is designed to solve the abundance of subpixel by using a wide range of spatial correlation effect and combine it with Conditional random field.In order to avoid the influence of pixel abundance error on sub-pixel mapping and eliminate noise artifacts in the results.The effectiveness of the method is proved by synthetic data experiments.