Application Of Region - Based Markov Random Field In High Resolution Remote Sensing Image Classification

With the successful launch of a large number of high spatial resolution remote sensing satellite at home and abroad in recent years, the acquisition of earth observation data has entered the era of high resolution satellite. But the ability of processing remote-sensing data in our China is far enough to satisfy the press for high resolution remote-sensing system. So the processing of high resolution remote-sensing data and the intelligent transformation of spatial information has become the important scientific problems in remote-sensing research field.High resolution remote-sensing use a very subtle way to observe the earth, and it can be more accurate to express the spatial relationship between them, and to provide a good condition and basis for the identify of ground object targets and the extraction of information about a scene. Then as the significantly improve of spatial resolution of remote-sensing images, ground object targets become more clear, and there is also many problems such as the complex and varied of ground object targets, the different spectra characteristics object with the same object,the different object with the same spectra characteristics, the different object with the same shape, the different shape with the same object, and the interference of complex background. Makes the traditional method for low and medium resolution remote sensing image processing is difficult to meet the demand of high resolution of remote-sensing application. In 2000,with the first commercialize software called eCognition-a software based on object-oriented appeared, makes the Geographic Object- based Image Analysis( also GEOBIA) get rapid development and the attention of remote sensing Image interpretation researchers. It represents the development trend of high resolution remote-sensing image information extraction.This paper will extend MRF model from the pixel level to region level, and the area will be used to model as the MRF model as the basic processing unit, to realize high resolution remote sensing image classification. In this paper, the research content mainly includes the following three aspects:(1) The basic knowledge of image classification based on MRF model and the algorithm of remote-sensing image segmentation. This chapter mainly elaborates MRF, Gibbs random field, Hammersley-Cliffor theorem; frequently-used label field models and feature field models; parameter eatimation algorithms and image segmentation optimization algorithms. The algorithm of remote-sensing image segmentation mainly includs the fundamental principles and typical image segmentation algorithms. In the end, segmentation and classification precision evaluation methods are introduced.(2) The high resolution remote-sensing image classification based on regional MRF model. The basic idea of region-based MRF model is described in detail including initial image segmentation algorithm, region adjacency graph(RAG) theory, region feature field modeling, region label field modeling, MAP estimation of region-based MRF, parameter estimation and the calculation process framework of region-based MRF model. In the final step, two types of images are used by this paper to verify region-based MRF model algorithm: synthetic remote sensing images with typical texture features and high-resolution remote sensing images. Conduct correlation analysis between parameters of pixel-based and region-based MRF models and classification results, and carry out contrast experiments with other mainstream classification methods(e.g. SVM, SVM with texture features, eCogniton, pixel-based MRF model etc.).(3) The high resolution remote-sensing image classification based on multi-scale MRF model. The classification of high resolution remote sensing image based on multi-scale MRF model, to expand the multi-scale MRF model based on pixel level to the multi-scale MRF model at the regional level. The build steps of the multi-scale MRF model at the regional level was introduced in detail: the segmentation of the high resolution remote sensing image used the mean shift method, get the initial segmentation result, and get the result of multi-scale segmentation according to the rules of merging. And then use the multi-scale region adjacent tree model, the association of inter-scale and intra-scale of the results of the modeling multi-scale region segmentation result; finally building the multi-scale MRF model on the multi-scale region adjacent tree model. At last,by the two types of images: to verify this region-based multi-scale MRF model algorithm by synthetic texture images and high resolution remote sensing images, and carry out contrast experiments with other mainstream classification methods(e.g. SVM with spectral features, SVM with texture features,, ecognition algorithm etc.)...