Research On Super-Resolution Mapping Technique For Remote Sensing Imagery

Abstarct:Although the remote sensing technology has been developing rapidly in recent years, there is still a big gap between the needs of high spatial resolution remote sensing image for the surface monitoring of environment and mapping and the acquisition ability of high resolution data such as repeat observation cycle is long, the coverage of images is small. Besides, the expensive price of high resolution image limits the needs of applications as well. Despite that the technology of mixed-pixels unmixing can offer some solutions to a certain extent, it can only ensure the composition proportion of each land cover class in mixed-pixels but fails to present the spatial location of land classes in the mixed-pixels. The monitoring of environmental in town with remote imagery is a weak link in environmental monitoring, and it is more difficult to achieve high resolution image in the region of town. Super-resolution mapping technology is a good method to predict the spatial position of land classes in mixed-pixels in sub-pixel spatial resolution which is of great significance for the environmental monitoring of towns.Using a synthetic multispectral image of some agricultural region in Netherlands as simulated data to conduct experiments, a part of Qiaokou town of Wangcheng country, Hunan province as study area. This thesis studied the following research through super resolution mapping technology:1. The feasibility study of existing super-resolution mapping methodsThe experiments of simulation images showed that those super-resolution mapping methods which were based on spatial correlation, pixels swapping, neural network, geo-statistics, landscape structure are feasible, despite that those methods of HNN, Markov random field (MRF) model, two histograms, BPNN and landscape structure had a poor ability to reconstruct boundary, even their results appeared "serrated" edge. But overall precisions were high and visual effects were good. It was provided that those methods were feasible on synthetic images.2. The suitability research of super-resolution mapping technology in study areaExperiments showed that the modified sub-pixel/pixel spatial attraction model (MSPSAM), mixed spatial attraction model (MSAM) and Pixel Swapping algorithm (PSA) had the best mapping results for study area. Combining the Linear Unmixing Model and the Maximum Spatial Dependence Model (SPMJLM) had the worst; except SPM_LM, the rest had higher overall accuracies than Support Vector Machine (SVM) hard classification, offered more details, and improved the spatial resolution of classification maps.3. The effect of the technology of mixed-pixels unmixing on the result of super-resolution mappingStudy showed that the technology of mixed-pixels unmixing had a tremendous impact on the results of super-resolution mapping. The results of those mixed-pixels unmixing methods which were based on full constrained least squares linear spectral mixing modeling (FCLS-LSMM) or Support Vector Machine (SVM) were unsatisfactory, but the result of SVM was better than FCLS-LSMM.