Research On Land Cover Change Detection Method Based On Spatial-spectral Feature Combination Using High Resolution Remote Sensing Image

Land cover change detection(LCCD)using very-high-resolution(VHR)remote sensing images is a technique for qualitatively and quantitatively discovering land cover change by analyzing bi-temporal or multi-temporal VHR images in the same geographical area.The technology has many practical applications,such as vegetation dynamic monitoring,agricultural resources assessment,nature disasters(earthquake,landslides and floods etc.)monitoring and prevention,land cover and land use,urban development planning,etc.Therefore,it is of great practical significance to carry out research on LCCD methods based on VHR remote sensing images.This paper focus on reducing the pseudo-change in the binary change detection map(BCDM)when VHR images are used for LCCD and improving the accuracy and reliability of LCCD,thereby providing reliable decision support for practical applications and needs.In view of this,the research idea of this paper is to take VHR images as the research object and take the image spatial-spectral feature extraction as the mainly research content to carry out susccessively based on different image analysis units(pixel level?object level?sematic level)for research on the LCCD.Therefore,the specific research in this paper includes the following three aspects:(1)Pixel level:This paper based on spatial-spectral feature of this adaptive region proposes three LCCD methods as follows:a)LCCD based on adaptive contextual information using bi-temporal remote sensing images.This approach defines an adaptive contextual mean feature(ACMF).Then,the change magnitude image(CMI)is generated by measuring the Euclidean distance between ACMFs.This method can effectively improve the detection accuracy because of the higher intra-class consistency in the CMI;b)Novel adaptive histogram trend(AHT)similarity approach for LCCD by using bitemporal VHR remote sensing images.AHT can effectively reduce the pseudo-change caused by the phenological difference through the spectral distribution trend,so that the inter-class separability between the changed area and unchanged area is enhanced,thereby increasing the detection accuracy;c)Based on the aforementioned research,this paper proposes a novel LCCD framework based on K-means clustering and adaptive majority voting using bitemporal remote sensing image.To utilize adaptive spatial information,this approach firstly applied K-means clustering to the local adaptive region in change magnitude image(CMI),and then the majority voting algorithm is used for decision the final class of the corresponding pixel.By this way,the detection accuracy is obviously improved.Notoably,the approach can be applied to other CMI to generate BCDM.The above three pixel-level approaches are tested by several real datasets and compared with other state-of-the-art LCCD methods.Experimental results show that the proposed three methods are relatively more advantageous in visual effects and quantitative accuracy.(2)Object level:The object-level method can not only smooth noises,but also usually better maintain the integrity of the BCDM,because the object can adaptively describe the borderline and shape of the ground target.Therefore,in this paper,two object-level approaches are proposed to further reduce the noise pixels in BCDM as follow:a)Multi-scale object histogram distance for LCCD using bitemporal VHR remote sensing images.This method utilizes the spatial contextual information through an object in adaptive and multi-scale manner.Moreover,a spectral frequency distribution histogram are constructed for each object,and the feature of the histogram is defined as arithmetic frequency mean feature(AFMF)to quantitatively described histogram of each object,which is designed to accurately obtain change information;b)Currently,LCCD(pre-processing)techniques focus on promoting a new change magnitude measuring or considering image spatial feature to acquire BCDM with less noise.To solve this problem,a post-processing approach is firstly proposed for refining the raw BCDM by a called object-based maximum expectation(OBEM)algorithm.The OBEM approach can further directly and concisely improve the detection accuracy based on the results of traditional methods.To demonstrate the effectiveness and performance of aforemetioned methods,several real datasets are used and compared with other state-of-the-art approaches in the experiments.Experimental results show that the proposed method has significant advantages in reducing noise and improving change detection accuracy.(3)Semantic level:Deep neural networks is an effective method to acquire higher-level semantic information of image for improving the performance of LCCD.In this context,this paper proposes a dual-path full convolutional network(DP-FCN)for quickly and automatically obtaining landslide inventory mapping(LIM)with bitemporal aerial remote sensing images.The DP-FCN model consists of two parallel deep feature extraction networks(DFE-networks)modules and a joint feature learning networks(JFL-networks)module.DFE-networks is firstly employed multiple convolutional layers for extracting image features,and then,low-dimensional and high-dimensional features are combined into final feature map for better capturing the semantic information of image.Finally,the JFL-network learns the potential semantic relationship between the image feature maps of bitemporal images to accurately obtain the landslide area.Experiments were conducted based on landslides on Lantau Island in Hong Kong and the results show that the DP-FCN model is effective and robust for automatically extracting LIM.