Spatial-Spectral Feature Representation And Learning For Change Detection In Remote Sensing Images

With the rapid development of social economy and technology,many realistic needs force people to change the Earth’s surface continuously.Based on the rapid development of remote sensing technologies such as satellite and aerial,a large amount of multi-source and multitemporal remote sensing images are constantly acquired.The change detection technology based on these remote sensing images has brought a boon for human to realize dynamic observation of the Earth’s surface and has been widely applied in important fields such as urban development planning,disaster prevention & mitigation and land resource management.Multi-source and multi-temporal remote sensing images contain rich temporal,spatial and spectral information,and are characterized by long span,high dimensionality and non-structure.However,how to effectively characterize the spatial-spectral features of remote sensing images has been the main problem in remote sensing image change detection.Therefore,this paper focuses on image spatial-spectral feature representation and learning,and carries out research on remote sensing image change detection methods and applications.Specifically,it includes the following aspects:(1)In image-level change detection,the spatial-spectral feature extraction often relies on a fixed-size sliding window,which is difficult to adapt to ground targets with different shapes and scales,thus limiting the performance of image-level change detection methods.To overcome this problem,this paper proposes an image-level change detection method with adaptive spatial-spectral feature representation.Based on the outlier discrimination capability of the box-plot principle,the method designs a parameter-free adaptive spatial neighborhood information extraction method to capture adaptive spatial contextual information with highly spectral homogeneous.Based on this,the corresponding adaptive spatial-spectral features are defined for each pixel to measure the change magnitude between pixel pairs in multitemporal remote sensing images.Experiments on real dataset show that the method achieves competitive results.(2)In the image-level change detection with multimodal remote sensing images,how to characterize the spatial-spectral features of multimodal remote sensing images with comparability is the main challenge in this type of change detection.To this end,this paper proposes an unsupervised multimodal change detection method based on the commonality spatial-spectral feature representation learning.The characteristics of this method are twofold: First,the method can obtain modality consistent reconstructed heterogeneous images by reconstructing the original multimodal images;Second,the commonality spatial-spectral features of multimodal remote sensing images can be learned gradually by minimizing the distance between the representative features of the reconstructed heterogeneous images.Finally,the change information between heterogeneous remote sensing images can be effectively obtained by directly comparing the commonality spatial-spectral features.The experimental results on six benchmark datasets show that the commonality spatial-spectral features based on multimodal remote sensing images can be effectively achieved and can improve the performance of multimodal change detection.(3)Buildings in remote sensing imagery present significant differences in spectra,shapes and scales,which leads to problems such as the low precision and the insufficient completeness of most building detection methods,and difficulty in balancing precision and completeness.To alleviate this problem,this paper proposes a building detection network based on global-local spatial-spectral feature learning.The method proposes a context-content fusion module to capture the long-range dependencies relationship and location information among buildings,so as to represent the global spatial-spectral features of buildings.Then,an edge residual refinement module is employed at the end of the network to efficiently utilize local spatialspectral features to refine the precision and completeness of building detection.Finally,a separated deep supervised strategy is introduced to explicitly guide the method toward obtaining higher precision or completeness.The experimental results show that this method can improve the overall building detection performance on three large-scale building detection benchmark datasets,and relatively balance the accuracy and completeness of building detection.(4)In building change detection,the performance of existing methods is still limited by various building features extraction and susceptibility to interference from other complex targets in dense urban scenes.To address this problem,this paper promotes a building change detection network based on multi-scale pyramid spatial-spectral feature learning and a cross-task transfer learning strategy based on building detection.In this method,a multi-scale pyramid building feature extractor is designed to effectively extract the multi-scale spatial spectral features of buildings.In addition,a cross-task building detection transfer learning strategy is introduced to pre-train the building change detection network,so that the building change detection network can focus on buildings,thereby effectively alleviating the interference caused by other ground objects.The experimental results show that the method can effectively mitigate the false pixels caused by other targets and significantly improve the change detection performance.(5)In recent years,the urban landscape of Wenzhou City,China,has undergone great changes,and the remote sensing image change detection technology can provide reliable geographic information data for the dynamic update of geographic conditions.Under this realistic demand,this paper systematically studies the performance of the current change detection methods in practical applications through the practical application case,i.e.,surface coverage change detection in the built-up area of Wenzhou City,China.In addition,to meet the application requirements of change detection in the built-up area of Wenzhou,this paper proposes a change detection network with integrated multi-dimensional spatial-spectral feature learning by combining the above-mentioned spatial-spectral feature learning modules.The experimental results on the real Wenzhou dataset demonstrate that most of the current change detection methods still have insufficient detection performance in practical applications,and the robustness needs to be further improved.Moreover,the change detection network with integrated multi-dimensional spatial-spectral feature learning in this paper can provide reliable change detection results for change detection application cases in built-up areas of Wenzhou.