Adaptive Conditional Random Field Classification Methods For High-resolution Remote Sensing Image

Compared with low-resolution and medium-resolution remote sensing images,remote sensing images with high spatial resolution can more accurately express the spatial relationship of features,especially in the aspects of clearer target feature contour and clearer geometric texture details.They are widely used in urban analysis,agricultural monitoring,disaster assessment,land survey and other fields.However,as the spatial resolution increases,the image information becomes highly detailed,and give rise to the scenarios of “same material with different spectrums” and “similar spectrum from different materials”.There are urgent problems to be solved.Traditional classification methods only use spectral information,and do not fully exploit the spatial information in the image,which limits the application requirements of classification.In the high-resolution remote sensing image method of object-oriented segmentation,the entire segmented object is taken as the analysis target of the image.The object is composed of pixels with similar characteristics,but the object-oriented classification result depends on the selection of the segmentation scale.Objects often exist in multiple scales,so it is difficult to select the optimal scale.Conditional random field is a model that can integrate the spectral and spatial context information of high spatial resolution images under a unified probability framework,and does not depend on the segmentation scale.Therefore,it becomes an effective method in the classification of high-resolution remote sensing images.However,there are still some problems with the existing classification method of conditional random fields:(1)When the existing model constructs multiple different potential functions,the parameters are mainly manually set,resulting in poor adaptability of the model;(2)At present,the model is mainly aimed at the fixed neighborhood around the pixel during model inference,and the spatial information of the image is not used at a larger scale.In response to the above problems,an adaptive conditional random field for highresolution remote sensing images was developed.The proposed model is combined with the spatial homogeneity information in the image,and make good use of advantage of the conditional random field model that can fuse multiple information such as image spectrum,texture,and position.The main research work and innovations are as follows:(1)Basic theoretical research on high-resolution remote sensing image classification model.This paper clarifies the basic theory of high spatial resolution remote sensing image for the classification application.And then,summarizes them into the followings: pixel-oriented classification methods,object-oriented classification methods,and random fields.Theoretical research based on conditional random field models was specifically carried out,including model construction,model training and model inference.(2)An adaptive conditional random field classification model based on spatial homogeneity is proposed.At present,the traditional classification method based on conditional random fields ignores the differences of multiple potential energy functions of pixels,and a single manual selection parameter leads to poor adaptability of the model.Therefore,an adaptive conditional random field model based on spatial homogeneity is designed in this paper.This model combines the spatiality of aerial images to achieve adaptive parameter control,effectively balancing the influence of one-element potential energy and two-element potential energy.In the framework of conditional random field models,this strategy of adjusting parameters for individual pixels can effectively fuse spatial context and positional relationship information.Experiments using Quick Bird,IKONOS and Worldview data,the experimental results verify the superiority of the algorithm compared with traditional algorithms.(3)A conditional random field classification model based on neighborhood adaptation is proposed.At present,more common algorithms based on conditional random fields all use fixed neighborhoods for model inference,which limits the model to infer the process more freely in a larger space.We adopted a shape-adaptive approach,set different neighborhoods for different pixels,and collected more spatial contextual information.The experimental results of Quick Bird and IKONOS data confirm the superiority of the algorithm compared with traditional algorithms.