Research On Sub-pixel Mapping Algorithms For Remote Sensing Image

With the rapid development of remote sensing technology nowadays,remote sensing images has been widely applied to many important fields,such as military observation,public security,traffic control,disaster prevention,geographical research,marine monitoring,agricultural statistics,forestry planning and so on.However,by certain factors,high resolution remote sensing images still play a less important role than the medium and low resolution remote sensing images In the process of production and application.Due to the limited spatial resolution,a large number of mixed pixels exist widely and inevitably in medium and low resolution remote sensing images.The problem of mixed pixels problem greatly limits remote sensing image information acquisition.So,it is the key problem point to solve the problem of mixed pixels in order to improve the spatial resolution of remote sensing image,which has attracted many scholars' attention for a long time.The issue of mixed pixels can be solved quantitatively by mixed pixel decomposition in some way.This technique can solve the spatial proportion of each land cover class within mixed pixels.But the technology can not get the specific distribution within the mixed pixels but only the spatial proportion.Sub-pixel mapping can make up the shortcoming of mixed pixel decomposition.Based on the fraction value by mixed pixel decomposition,sub-pixel mapping divides mixed pixels into sub-pixels,which estimates the spatial distribution within mixed pixels on higher spatial scale.Sub-pixel mapping and its relevant issues have been researched in this paper.Studies on sub-pixel mapping are mainly conducted on the following three aspects:1.Referring to the fast sub-pixel mapping algorithm based on interpolation,this paper proposed a new sub-pixel mapping method based on image self-similarity,a kind of super-resolution reconstruction technique.The proposed method learned the structure similarity and corresponding relationship between the image itself and its degraded image,built dictionary entries of low-resolution block and corresponding high-resolution block,matched similar blocks to estimate the high-resolution block of each low-resolution block and stitched all the blocks.After class allocation on high-resolution fraction images,sub-pixel mapping can be realized.2.An improved new BP network model is proposed after careful study on existing sub-pixel mapping methods based on BP neural network.Vectors composed of the attribute value of central mixed pixel and its 8 neighbor pixels are taken as input of traditional BP network models.This proposed method changed the input and output of the traditional BP neural network,improved BP network model through the new relationship:taking partial block of low-resolution fraction images as network input,and sub-pixel block corresponding to the low-resolution block as the output.The new model considered more about spatial correlation hypothesis and local structure information and showed better mapping accuracy.3.A sub-pixel mapping method based on improved self-correlation BP network is proposed.This improved method takes multiple shifted images as constrains and modifies the original model by changing the way how Moran'I works.Allocate land cover classes initially on the soft attributes of network output.List all the possible distribution by constrains of shifted images and calculate the Moran' I of each possible distribution.The distribution of maximum Moran' I is then considered as final distribution.Experimental results show that the proposed method has higher computational efficiency and mapping accuracy compared with the classical self-correlation BP network algorithm.In this paper,the above three methods are mainly studied.Method based on image self-similarity takes advantage of the idea of super resolution reconstruction,makes fully use of the image spatial information;The new method based on BP network refers image self-similarity based method,changes the input and output of classical BP network,which fits the spatial correlation hypothesis better and improves the mapping accuracy;Combined with multiple shifted images and spatial self-correlation,the sub-pixel mapping method based on improved self-correlation BP network improves the process of land cover class allocation,reduces the uncertainty of sub-pixel mapping.Different emphases are laid on three methods.