Research On GA And MRF Based Sub-pixel Mapping

Hyperspectral remote sensing is the multi-dimensional information obtaining technology,and it not only could obtain the two dimensional spatial information which is used to describe the objective distribution, but also could get the one dimensional spectral information. The spectral resolution of hyperspectral remote sensing image is very high,and people’s recognition ability goes deeper along with the increasing of the spectral information, but the spatial resolution of the hyperspectral remote sensing image is still very low and mixed pixels widely existed in remote sensing image, which has brought great difficulty in visual inspection and post-application. Referring to the mixed pixel processing, hard classification will lose a large amount of land cover information, because of this, some researchers proposed the soft classification methods. More specifically, it includes endmember extraction,abundance estimation and sub pixel mapping. Endmember extraction could extract the pure endmember in the hyperspectral image, and abundance estimation techniques have been conducted to calculate the proportion occupied by each land cover class in the mixed pixel,and sub pixel mapping technology could predicted the most probable spatial distribution.In this paper, some of the important issues in sub pixel mapping have been discussed.The concrete work is shown as follows:Firstly, some sub pixel mapping algorithms, such as SPSAM, pixel swapping algorithm(PSA) and the like, are studied. SPSAM assigned the land cover class to the sub pixel directly,and the attractive value is very rough, due to these two reasons, it will generate a lot of isolated pixels in the final results. Although PSA has very high iterative speed, it is very sensitive to the noise and the initial distribution.Then, genetic algorithm which could be applied in the sub pixel mapping has been studied also. The crossover operator in GA selects the gene at random, which makes the iterative efficiency is very low, and the final sub pixel mapping accuracy is not high also. In this paper, a novel method, which is called Modified Genetic Algorithm (MGA), is proposed to realize SPM, and it not only combined the advantage of population, but also combined advantages of PSA, so it could further improve the iterative efficiency.Lastly, the aforementioned sub pixel mapping algorithms put the fraction images as the input, and the existing unmixing algorithms are difficult to meet the accuracy requirements,which makes the error overlapped in the final sub pixel mapping results. Aiming to these demerits, this paper first describes the Markov Random Field (MRF) which is could be used in the sub pixel mapping field. Due to MRF could be combine the spatial and spectral information at the same time, this paper further describes the multiple spectral constraints based sub pixel mapping. Although it could improve the accuracy to some extent, it could not consider the spatial information of SSRSI, which makes the accuracy still very low. So this paper proposed a multiple spatial constraints and multiple spectral constraints based sub pixel mapping algorithm, and it could further improve the SPM accuracy.