Conditional Random Fields For High Resolution Remote Sensing Image Classification

With the rapid development of earth observation technology,the spatial resolution of remote sensing images has been greatly improved,and has been able to reach the sub-meter level.High-resolution remote sensing images can clearly express the boundary of land-cover classes,and provide rich spatial information,which enables detailed image classification.However,with the increase of spatial resolution,the image information is highly detailed,and brings us the scenarios of“same material with different spectra”and"similar spectra from different materials".This problem weakens the spectral separability of land-cover classes,which poses challenges to the classification of high-resolution remote sensing images.In order to exploit the spatial advantage of high-resolution remote sensing images,recently some scholars have proposed the classification methods integrating image spatial information,which mainly contains image classification of spatial structure features,object-oriented classification,and random field method.For the image classification of spatial structure features and object-oriented classification,the classification accuracy can is improved by considering the image spatial information,but they respectively face the problems that the feature dimension is high and the best segmentation scale are difficult to choose.The random field model uses the undirected graph structure to express high-resolution remote sensing images,which can model the spatial context information in a unified probability framework,and has the advantages of clear comprehensibility,unified framework and strong modeling ability.As a typical random field model,the Markov random field can model the context information in labeling field.However,for computational tractability,it is assumed that the observed information is conditionally independent,limiting the flexibility of the use of spatial information.The conditional random fields formally models the posteriori probability for the classification problem,which can flexibly construct the spatial context information under the unified probability framework,and is more suitable for remote sensing image classification.However,when the model are applied to high-resolution remote sensing images with complex distribution and spectral variation,the conditional random field model still faces the problem that the large-scale spatial context information description ability is limited,the inference method is difficult to take into account the efficiency and expansibility,and the application scope is limited to the pixel level classification.To overcome the problems of conditional random fields(CRF)for high-resolution remote sensing image classification,this thesis develops the classification methods based on conditional random field model for high-resolution remote sensing images from three aspects of model construction,model inference,and model application.The main contents of this thesis are summarized as follows:(1)This thesis systematically introduces the basic theory of conditional random field model,and analyzes the characteristics of high-resolution remote sensing images and the difficulties of high-resolution remote sensing image classification.In addition,the thesis also introduces the commonly used algorithms of high-resolution remote sensing image classification.(2)In the aspect of model construction,object-oriented conditional random field classification algorithm and higher-order conditional random field classification algorithm integrating spectral-spatial-location cues are proposed for high-resolution remote sensing images.For the traditional CRE,the existing potential function does not fully consider the characteristics of high-resolution remote sensing image,and lacks large-scale spatial interaction modeling ability,so that the model may have an oversmooth performance.In this thesis,object-oriented conditional random field classification algorithm is designed.The CRF absorbs the advantage of object-oriented method to consider large-scale spatial context information based on the object unit.For the limitation of the pairwise CRF in the modeling of large scale spatial context information,the high-order conditional random field classification method is developed.The high-order potentials is calculated to obtain large scale spatial contextual information as the spatial location cues.The given algorithm integrates spectral,spatial contextual,and spatial location cues within a CRF framework to provide complementary information from varying perspectives,so that it can address the common problem of spectral variability in remote sensing images.(3)In the aspect of model inference,a conditional random field classification algorithm based on cooperative game theory is designed for high-resolution remote sensing images.The traditional optimization is difficult to take into account both efficiency and accuracy.Therefore,the given algorithm refers to game theory.The local expected energy based on the mixed strategy can be calculated independently for each pixel to consider the uncertainty of the labels and the local spatial interaction.The merge rule can be established based on the local expected energy to form coalitions,which means that the pixels in each coalition share their information in a cooperative game.A simple cooperative game,majority game,is then used to make the coalition decision to obtain the labels of the pixels.(4)In the aspect of model application,a sub-pixel mapping algorithm based on conditional random fields is designed.At the sub-pixel level,the spatial distribution of sub-pixel is emphasized.The sub-pixel mapping algorithm based on conditional random fields designs the potential functions to integrate the local spatial prior at the fine scale and the downscaled coarse fraction at the coarse scale,which can make full use of the input fraction image and the spatial dependence of different scales to obtain more detailed land-cover distribution information.Considering the function of abundance to determine the spatial distribution of the sub-pixels and fraction errors,the probability class determination strategy is designed to take the abundance constraint as a soft constraint to utilize the input fraction image.(5)High-resolution remote sensing image classification framework based on conditional random field is built.Combined with high-resolution remote sensing image classification method based on conditional random field proposed from multiple perspectives,the high-resolution remote sensing image classification framework based on conditional random field is built.The classification framework can be used for large-scale high-resolution remote sensing image classification.This thesis conducts the research of conditional random fields in model construction,model inference,and model application for high-resolution remote sensing image classification.To fully explore the spatial contextual information of high-resolution remote sensing images,a series of conditional random field classification algorithms under different conditions are designed,which can lead to detailed classification.