Research On Change Detection Methods Based On Adaptation Neighborhood Scale Using High Resolution Remote Sensing Images

High-resolution remote sensing image change detection technology is a technique that uses high spatial resolution remote sensing images(referred to as "Very-High-Resolution images")of the same area at different times to discover and identify surface change information.Currently,it has been widely used in natural disaster assessment,land use change investigation,national defense and security,among other areas.Due to the different shapes,sizes,and types of objects contained in high-resolution images,traditional fixed and single scales are difficult to effectively capture the complete features of objects.This leads to inconsistent object scales in high-resolution images,and change detection applications typically face the following challenges:1)traditional single-scale features are difficult to accurately measure the magnitude of changes between different types of objects;2)the low-level features of objects in two high-resolution images are easily affected by external factors,resulting in incomplete feature scale extraction and insufficient expression of object information;and 3)the ability to express deep semantic scale information is weak in data model training.Based on the above problems,this paper uses common highresolution images,combined with image spatial-spectral information and deep semantic features to carry out change detection research,aiming to break through the difficulties and challenges in change detection.The main work of this paper focuses on adaptive neighborhood scale feature mining to achieve higher accuracy in high-resolution image change detection.The specific research is summarized as follows:(1)For the problem that a single scale feature in high-resolution imagery is difficult to accurately measure the magnitude of changes between various types of objects,this paper proposes two methods:a)A detection method based on adaptive scale shape similarity.This method defines an adaptive shape distance and uses direction pixel sparsity to express the difference between the dual-time image features;b)A detection method driven by spatial context scale information.This method defines a new adaptive expansion algorithm that reduces parameter constraints while improving the automation degree of the algorithm.Both methods are verified through multiple data sets and compared with other methods.Experimental results show that the proposed methods can effectively take into account the scale information of obj ects,avoid isolated noise and the influence of "pseudo-changes" in objects,and improve the separability of change and unchanged categories.(2)For the problem that the low-level features of objects in two-phase high-resolution imagery are susceptible to external factors,leading to incomplete feature scale extraction and insufficient information representation,this article proposes a detection method based on the visual spatial-spectral attention mechanism network.The method constructs a novel feature fusion network structure that uses the change magnitude feature to guide and combine the spatialspectral attention module to learn the intrinsic relationship between the features of the two-phase images and express the feature information difference.Experiments were conducted on several datasets and compared with similar methods.The results show that the proposed method can achieve precise positioning of the change regions in a fast manner.(3)For the problem of weak expression ability of deep semantic scale information in data model training,a detection method based on multi-scale deep feature fusion network is proposed.This method designs a multi-scale adaptive convolution block and re-optimizes the loss function.In the encoding stage,this method extracts features at various hierarchical levels of the two-phase images.In the decoding stage,the feature maps at different hierarchical levels are fused into a deep supervision network with different branches to reconstruct multi-level change maps.Experiments are conducted on datasets from different application scenarios and compared with other similar research methods.The experimental results show that the proposed method model has better effectiveness and robustness.