Change Information Extraction For High Resolution Images Based On Conditional Random Fields

Multi-temporal remote sensing images have the advantage of providing long-time and continuous observation for land cover.By using remote sensing image change detection technology,the change information in multi-temporal remote sensing images can be extracted,which can qualitatively and quantitatively reflect the change of land cover,and provide technical support for land survey,urban planning,ecosystem monitoring,disaster monitoring and assessment,military reconnaissance and other applications.With the launch of a series of high resolution satellites,the spatial resolution of multi-temporal remote sensing has greatly increased,which provides a data basis for fine change detection.However,due to the inherent spectral variability of high spatial resolution images,i.e.the increase of intra-class variance and the decrease of inter-class variance,the change detection results are more susceptible to the influence of different temporal image spectra and geometric quality differences,which results in the difficulty of distinguishing the pixels with different change types,and hinders the application of multi-temporal high resolution images in practical production.Recently,spatial-spectral feature extraction and object-oriented methods are commonly used to extract the spatial information in remote sensing images.However,these methods still have some limitations,such as the difficulty in feature selection and scale selection.It is difficult to take into account the local-global characteristics of images and have poor robustness to different research areas and different data source algorithms.Random field,as a probabilistic undirected graphical model,has the ability of modeling the contextual global statistical characteristics of images in a unified probabilistic framework through the construction of local potential function.It has attracted much attention in the field of image processing.Markov random field is a traditional random field model,which can effectively model the spatial contextual information in the labeled field.It has been widely used in low and medium resolution optical and SAR image change detection.However,as a generative model,the strong independence assumption in the observed field limits the application of the model in high spatial resolution image change detection with more complex spectral and spatial change patterns.As a discriminative probability graphical model,conditional random field has the advantage of flexible modeling compared with Markov random field,and is more suitable for processing high spatial resolution image change information extraction.However,there is no systematic random field model for high spatial resolution remote sensing image change detection.The following problems still exist when applying conditional random field to change detection for high spatial resolution remote sensing images: 1)the ability of unary potential to discriminate the changed/unchanged samples is weak;2)the pairwise potential cannot describe the fine structural information of high resolution images;3)the large-scale spatial contextual information cannot be taken into account;4)the model does not have the ability of multi-type change analysis.Based on the above mentioned questions,this paper systematically studies the application of conditional random fields in high spatial resolution remote sensing image change detection,and creatively proposes a high resolution remote sensing image change detection framework based on conditional random fields.The main contents and innovations of this paper include:(1)The state-of-the-art optical remote sensing image change detection methods are carefully reviewed,and the main factors affecting the accuracy of high spatial resolution image change detection are originally summarized.(2)Aiming at the problem of weak discriminative ability of conditional random field model to change/unchanged samples,a multi-feature fusion conditional random field model is proposed.The spectral and spatial information of pixels are considered simultaneously in the modeling of unary potential to improve the accuracy of model initialization to get a better result in the inference step.(3)In the exploration of spatial structural information,the traditional pairwise conditional random field uses isotropic pairwise potential,which cannot take into account the shape and orientation of the local structure of the image.In this paper,a structural conditional random field is proposed,which uses pairwise conditional random field model with different shape and direction of pairwise potential structure to give the complete description of high resolution image change information,and takes into account the size and direction characteristics of image local structure features to extract the optimal feature expression of image local structure.(4)For large-scale spatial contextual information extraction,a hybrid conditional random field model is proposed,which combines object-oriented method.The commonly used pairwise conditional random field model constructs the spatial context information between neighboring pixels through pairwise potential,which cannot effectively extract the spatial information of image objects,making the model ineffective in discriminating the changed types at object scale and boundary is not well preserved.For this reason,the traditional conditional random field model is combined with object-oriented method by constructing object potential,which makes the model get the ability of pixel-neighborhood-object-global change information mining.(5)In the aspect of change analysis,a conditional random field model based on label space is proposed.The traditional conditional random field model relies on the pixel discriminative information provided by the observed image in the process of modeling.When facing different data source,the traditional change detection method is difficult to recognize the change type effectively.Therefore,a conditional random field change analysis model in the label space is proposed to transform image data into label space,so that different image data source and thematic data can be compared under a unified framework.It is also advantageous to construct conditional random field potentials based on local statistical features of super-pixels,so that the random field model takes into account both spatial detail preservation and multi-type analysis ability.Aiming at the application of multi-temporal and high spatial resolution remote sensing image change detection,this paper systematically summarizes and analyses the factors affecting the change detection results,carries out targeted research on conditional random field model for high spatial resolution image change detection,and constructs a prototype system of high resolution image change detection based on conditional random field,which can be provided for application departments.