Research On The Change Detection Method Of High-resolution Remote Sensing Image With Multi-scale And Multi-dimensional Features

The remote sensing image change detection is an important technical tool for monitoring the dynamic change of global surface landscapes.It has applications in a wide range of fields,including land survey,urban research,ecological environment monitoring,and disaster detection.As the ability to acquire a multi-temporal high spatial resolution remote sensing image(referred to as high-resolution images)increases,it provides massive data support for the accurate extraction of remote sensing image change detection information.In contrast with low-and medium-resolution remote sensing images,high resolution images have new features,such as spectral heterogeneity of homogeneous areas,detailed surface information,and spatial multi-scale geographic information,which present new challenges for remote sensing image change detection.Specifically,(1)with the increase of spatial resolution,high resolution images have a series of characteristics such as fine spectral features,spatial geometric features,complex spatial topological relationships of features,and spectral heterogeneity of homogeneous regions,which make it difficult for the change detection system of traditional low and medium resolution remote sensing images to meet the needs of high-resolution image change detection.(2)The highly detailed surface information from high resolution imagery causes an increase in variance within features and a decrease in variance between feature classes.This phenomenon makes it difficult for pixel-based change detection methods to solve the difficulties of "same object with different spectrum" and "different object with same spectrum" caused by spectral interpretation.(3)The surface landscape is manifested as multi-scale unity in the ground observation of high-resolution images,and it is difficult to decipher the spatial multi-scale characteristics of the surface landscape in high-resolution images using a single observation scale.A smaller observation scale leads to discrete large-scale features and a larger observation scale leads to an over-integration of small-scale features.Therefore,in order to improve the accuracy of change detection,it is especially important to use more rational and effective work procedures.The new features of high-resolution images,such as spectral heterogeneity in homogeneous regions,detailed surface information,and spatial multi-scale of geographical information,pose challenges to change detection.In this paper,a multi-level interpretation model of the "pixel-object-scene" is presented to exploit the rich "spectral-spatial" deep multi-dimensional features in high resolution images and enhance the semantic consistency of multi-scale objects.And this paper focuses on the application of multidimensional features of high-resolution images in change detection under multi-scale objects.Finally,we analyze the advantages and disadvantages of the three proposed change detection algorithms in different application scenarios to provide technical and theoretical references for efficient and fast,intelligent change detection.The research contents and results of this paper are as follows.Change detection of high-resolution images with joint pixels and multi-scale objects is presented.First,this study utilizes classical evidence theory to fuse pixel evidence with object evidence to provide clear boundary information for high-resolution image change detection.Second,when using traditional evidence theory for change detection studies,there are shortcomings of inheriting uncertainties in the parent segmentation scale to the child segmentation scale and the choice of object segmentation algorithms is limited by the parent-child segmentation.In this paper,to avoid the incremental error in the "parent-child" segmentation analysis,we improve the performance of the multi-scale object change detection algorithm by eliminating the influence of parent scale on child scale segmentation through a scale-driven majority voting strategy.Finally,a comparison experiment using GF-2 images from four different regions is conducted to verify the validity of this method.Object-based change detection based on multidimensional feature visual saliency detection is presented.Traditional visual saliency detection uses grayscale features of differential images to characterize differences in geographic elements,making it difficult to fully utilize the unique spectral and spatial structure features of high-resolution images.In this paper,the expression of the visual perceptibility of remote sensing images in spectral-spatial features is categorized into local low-level features,global low-level features,and high-level features,to reveal the deep multidimensional features of remote sensing images in the process of visual saliency detection.To solve the problem that the pixel-by-pixel method loses the contextual local information,change detection is accomplished by comparing the differences between an object and all objects in its k-level neighborhood.At the same time,the multi-scale weighted fusion strategy is used to fuse the significance intensity of different scales,which solves the over-segmentation and under-segmentation problems that occur in the single-scale change detection results.Comparative experiments on multi-source remote sensing image datasets showed that this method could preserve deep-level information of images while resisting noise,and achieved better experimental results in high-resolution images compared with other methods.Convolutional Neural Networks(CNN)with automatic annotation for change detection driven by multi-scale objects is presented.After classification,the comparison can meet the need of specifying feature change types in heterogenous remote sensing image change detection.Based on the experience of successful application of deep learning to feature type recognition,deep learning has become one of the main research elements in heterogenous remote sensing image change detection.In response to the lack of large-scale,multi-level remote sensing interpretation datasets in deep learning,time-consuming and laborious manual acquisition of datasets,and single source of datasets,the study proposes a technical process to establish efficient and high-quality datasets using vector data for automated annotation of samples,and provides the possibility of continued expansion in terms of diversity,richness,and multi-scale of datasets.For multi-scale datasets,the study establishes a multiscale object-based CNN to realize the identification of land use types in large-scene remote sensing images by fusing patch-based CNN,object-based voting strategy and multi-level majority voting strategy.The improved method was applied to SPOT-6,GF-2,and ALOS high-resolution images for feature type identification,and the change detection of comparison after classification of heterogenous remote sensing images was achieved.The effectiveness and robustness of the method for heterogeneous remote sensing image change detection were verified by extensive experiments.