Research On Sub-Pixel Mapping And Its Related Techniques For Remote Sensing Imagery

Using remote sensing imagery, human’s view has extended from land to space, whichhas greatly improved people’s earth observation ability in many aspects. However, due to thenature of the real landscape and the data acquisition process, mixed pixels exist widely inremote sensing images. These pixels contain multiple land cover classes. It is certainlyinappropriate to allocate these mixed pixels to just one class as much information will be lost.The mixed pixel problem causes much trouble in land cover information extraction.Soft classification has been developed to estimate the proportion of each land cover classin mixed pixels. However, it fails to present the specific spatial location of land classes in themixed pixels. Sub-pixel mapping or super-resolution mapping has been developed to predictthe spatial location of land cover classes at finer (sub-pixel) spatial resolution. As a potentialtechnique to increase the spatial resolution of land cover map, sub-pixel mapping has been ahot issue in the field of remote sensing for years.The work of this thesis mainly focuses on sub-pixel mapping. Also, some relatedresearch has been done on soft classification, a pre-processing step of sub-pixel mapping. Thedetailed work of the thesis is given below:1、A fast geometric estimation method for linear spectral mixing modeling (LSMM) wasproposed. The traditionally iterative process for solving LSMM is of heavy computationalburden. Recently developed geometric analysis methods for LSMM have provided a new wayto reduce the complexity of LSMM solving. The unmixing results, however, are not in linewith the fully constrained (i.e. non-negative and sum to one constraint) least squares (FCLS)requirements. Aiming at this issue, a new geometric unmixing method was constructed tocompletely meet the FCLS requirements. The theoretical analysis and experiments showedthat the proposed method was of very low complexity and can obtain the theoretically optimalsolution as well.2、A least squares support vector machine (LSSVM) based sub-pixel mapping methodwas proposed for land cover class with linear features. The back-propagation neural network(BPNN) based sub-pixel mapping method requires lengthy training and is susceptible to thelocal optima. In addition, many training samples are required to train BPNN. To eliminate the dependence on prior information from land cover class with linear features, a geometricmethod for artificially synthesizing training samples was proposed according to the linearfeatures. Moreover, sub-pixel mapping was conducted by using LSSVM, an efficient learningalgorithm with high generalization ability. Experiments showed that the method combiningLSSVM and synthetic training samples is feasible and compared with BPNN, the trainingprocess of it was much faster and can attain sub-pixel mapping results with higher accuracy.3、A new sub-pixel mapping algorithm was proposed based on modified sub-pixel/pixelspatial attraction model (MSPSAM). Using the idea of differential and integral, MSPSAMconsiders the spatial distribution of each sub-pixel within neighbor pixels while calculatingthe spatial attractions for sub-pixels within the centre pixel. Then the attractions are used todetermine the class values of these sub-pixels. Experiments results showed the effectivenessof MSPSAM.4、A mixed spatial attraction model (MSAM) for sub-pixel mapping was proposed.Based on MSPSAM, MASM takes the dependence among sub-pixels within mixed pixel intoaccount and integrates the spatial attractions both within and between pixels. According to theexpression of the MSAM maximumising the spatial attraction, the genetic algorithm wasemployed to search the optimum solution and obtain the sub-pixel mapping results.Experiments showed MSPSAM method can provide high accuracy sub-pixel mapping results.5、A new Markov random field (MRF) based sub-pixel mapping model was proposed byusing multiple constraints. In nature, the sub-pixel mapping problem is an under-determinedproblem. The effectiveness of conventional MRF based sub-pixel mapping approach islimited by the single spatial and spectral constraints. In this thesis, the spectral informationfrom sub-pixel shifted remote sensing images (SSRSI) was incorporated into the likelihoodenergy function of MRF to provide multiple spectral constraints. Experiments showed that theproposed sub-pixel mapping method can provide high accuracy results.