Research On Sub-pixel Mapping For Remote Sensing Image

With the rapid development of remote sensing technology, remote sensing image has been widely used in environmental or resource management, natural disaster monitoring,agriculture or vegetation planning, public security and other technical areas. However, the existence of mixed pixels are inevitable, which are influenced by the environmental factors,the resolution of sensor and other factors in the data acquisition process. The existence of mixed pixels limits the spatial resolution of remote sensing images. This has brought great difficulties in acquiring of land cover information. Therefore, how to improve the spatial resolution of remote sensing images has become one of the heated problems.Spectral un-mixing technology has been developed to get the proportion of each land cover class in mixed pixels. However, it fails to determine the specific spatial distribution of each land classes for each land cover class in mixed pixels. Sub-pixel mapping is an efficient technology to determine the specific spatial distribution in mixed pixels. It makes the distribution information displayed at a finer spatial resolution. Based on this information, this paper careful study and summarize the existing research results firstly, then made a deep study on sub-pixel mapping for remote sensing images. The main research content is described as follows:1. A new sub-pixel mapping method based on the sub-pixel/sub-pixel spatial attraction model (SSSAM) and the analysis of model parameters. After careful analysis and study weight distance function has been found to be a key for describing spatial correlation.Different kinds of weight distance function can interpret the spatial correlation from different perspectives. It affects the performance of sub-pixel mapping to a large extent. In this paper,three weight distance functions including distance bottom model, exponential covariance model and gaussian model were used into SSSAM respectively. In order to choose out the optimum weight distance function, two groups of experiments were used to test the performance of the mapping methods with the three different weight distance function.Experiments show that distance bottom model has the worst accuracy. Gaussian model has a higher accuracy than exponential covariance model, but gaussian model is more sensitive to variable parameter.2. A new sub-pixel mapping method of remote sensing image based on cubic convolution interpolation is proposed. Sub-pixel mapping can be described as the following two steps: First, the probabilities of class occurrence at each sub-pixel are estimated by super-resolution technology. Second, the thematic class is allocated to each sub-pixel,according to those probabilities and constraints from spectral un-mixing. This process can be seen as a hard classification on per sub-pixel. Based on the above, the super-resolution result of low-resolution image obtained by spectral un-mixing is achieved by the cubic convolution interpolation algorithm firstly. After hard classification on per sub-pixel, the classification figures on sub-pixel basis were obtained. Experiments show this proposed sub-pixel mapping method has faster speed and higher accuracy, and is simpler and easier to implement.3. A new sub-pixel mapping method for remote sensing image based on hybrid interpolation is proposed. Although the mapping method using traditional interpolation algorithms is feasible, the edge blurring of interpolation algorithms itself exist inevitably. To avoid affecting on the performance of sub-pixel mapping, this paper presents a new interpolation algorithm, named dual interpolation. Combine dual interpolation (DI) with bilinear interpolation (BI), marked as BI-DI, and dual interpolation with inverse distance weighted interpolation (IDWI), marked as IDWI-DI, two kinds of hybrid interpolation are obtained. The hybrid interpolation algorithms are used in the first step of the sub-pixel mapping to predict the soft class value of each sub-pixel, while class allocation is employed in the second step to estimate the hard class for each sub-pixel according the constraints from spectral un-mixing?Experiments show that, compared with the methods using only one interpolation algorithm, the proposed sub-pixel mapping methods can retain image edge feature better and produce higher accuracy.In this paper, three sub-pixel mapping methods were studied. First, different weight distance functions were used into SSSAM respectively. Two groups of experiment are used to choose out the optimum weight distance function. Then, traditional interpolation algorithms were applied in sub-pixel mapping. It developed a new approach for sub-pixel mapping. This method can achieve good performance without training sample and iterations. The method with hybrid interpolation is proposed on this basis of this method. The method uses dual characteristic between high-resolution image and the low-resolution image to overcome the edge blurring in traditional interpolation algorithm. To further improve the precision of sub-pixel mapping method. All of the three methods can improve the performance from different perspectives. It has important significance for the applications of remote sensing images.