Sub-pixel Mapping Theory Considering Spatial Characteristic For Remote Sensing Imagery

Due to the limitation of the spatial resolution and the complicated scenes, mixed pixels are common in remotely sensed imagery. The mixed pixel problem brings difficulties to analyze the remotely sensed images quantitatively. Spectral unmixing is an efficient technique to solve the mixed pixel problem by building the spectral mixing model to obtain the abundance fraction of different land covers in the mixed pixels. However, spectral unmixing technique does not obtain the sub-pixel spatial attribution of land covers in mixed pixel, and the sub-pixel spatial information will be lost. It has become very important to do sub-pixel analysis by determining the spatial distribution of different land covers in mixed pixel.The sub-pixel mapping technique can obtain a fine-resolution map of class labels from coarser spectrally unmixed fraction image by dividing a pixel into sub-pixels and assigning these sub-pixels to different classes. Most of traditional sub-pixel mapping methods were based on the fraction image generated from single remotely sensed image, and the sub-pixel mapping results were obtained by maximizing the spatial correlation. There are the following problems for traditional sub-pixel mapping methods.(1) The spatial structure characteristic at sub-pixel level, such as the linear feature in mixed pixel, was not considered, resulting in the loss of spatial features.(2) The accuracy of sub-pixel mapping was limited due to the insufficient information of single image, and it can be improved by incorporating auxiliary spatial information at sub-pixel level.(3) Spectral unmixing and sub-pixel mapping were implemented separately. Both two procedures can not utilize the spatial information and they should be unified.To solve these problems, this thesis focuses on the sub-pixel mapping theory and methods considering spatial characteristic for remote sensing imagery. The main research work and the corresponding contributions of this thesis are as following:(1) The thesis summarizes systematically the current main models and algorithms of sub-pixel mapping. The principle and traditional methods of sub-pixel mapping were described in detail and the application of sub-pixel mapping were discussed and summarized.(2) A multi-agent system based adaptive sub-pixel mapping method is proposed to reconstruct different spatial features in mixed pixels, such as the linear sub-pixel features. Most traditional sub-pixel mapping approaches ignore the different inherent structures in mixed pixels and consider the mixed pixels as an identical type, for example, boundary features, leading to incomplete and inaccurate results. In the proposed sub-pixel mapping framework, three kinds of agents, namely feature detection agents, sub-pixel mapping agents and decision agents, are designed to solve the sub-pixel mapping problem. Feature detection agents are first created to detect the different structures in mixed pixels, then the sub-pixel mapping agents are used to reconstruct different features and the decision agents are introduced to obtain the optimal sub-pixel mapping result lastly. The experimental results indicate that the proposed algorithm outperforms the other two traditional sub-pixel mapping algorithms in reconstructing the different structures in mixed pixels.(3) A spatial-temporal attraction model is proposed to incorporate auxiliary spatial information at sub-pixel level of multiple shifted images. Traditional sub-pixel mapping algorithms only utilize a low-resolution image, the information of which is not enough to obtain a high-resolution land-cover map. The accuracy of sub-pixel mapping can be improved by incorporating auxiliary datasets, such as multiple shifted images in the same area, to provide more sub-pixel land-cover information. By extending the traditional spatial attraction model to the proposed spatial-temporal attraction model, the impacts of base image and auxiliary images are quantified as spatial attraction value and temporal attraction value. Then the two values can be combined given different weights and a class determination strategy is utilized to obtain the final sub-pixel mapping result. The experimental results show that the proposed model is successful in incorporating spatial information of multiple shifted images and reducing the impact of unmixing error to improve the accuracy of sub-pixel mapping result.(4) A sub-pixel mapping framework based on a maximum a posteriori (MAP) model is proposed by utilizing the complementary information of multiple shifted images to decrease theimpact of unmix error. By constructing the multiple shifted images observation model, all low-resolution fraction images can be taken into consideration equally. Then the MAP is utilized to obain the energy function by incorporating the spatial distribution prior of land covers. Lastly, the high-resolution MAP results for all land covers should be integrated to obtain the final sub-pixel mapping result. Experimental results demonstrated that the proposed approach outperforms the traditional sub-pixel mapping algorithms, and hence provides an effective option to improve the accuracy of sub-pixel mapping for remotely sensed imagery.(5) A unified sub-pixel mapping model integrating spectral unmixing is firstly proposed to utilize spectral and spatial information simultaneously. Traditional sub-pixel mapping methods were imposed on the fraction image which was generated with spectral unmixing techniques. Obviously, the capability of sub-pixel mapping was limited by the accuracy of the obtained fraction image. Moreover, the implementation of spectral unmixing was thought to be independent of sub-pixel mapping while they are correlated. A unified sub-pixel mapping model is firstly built and then the Lagrangian method is used to generate the energy function by incorporating spatial distribution prior. Lastly, the gradient descent method is utilized to obtain the optimal sub-pixel mapping result. Experimental results show that the proposed method is an efficient sub-pixel mapping technique for improving accuracy compared with traditional fraction image based sub-pixelmapping methods.