Research On Remote Sensing Image Change Detection Method Based On Spatial-temporal Semantic Information

Remote Sensing Image Change Detection aims to process multi-temporal remote sensing images to mark out the changing building entity areas,and has a wide range of application scenarios in engineering practice,such as urban construction planning,disaster assessment,unauthorized building detection,etc.In recent years,with the explosive growth of the number of remote sensing satellites and the popularity of high-resolution remote sensing satellites,the high-spatial and high-temporal resolution of remote sensing satellite images bring unprecedented challenges to the change detection task.Remote sensing image change detection methods based on deep learning have become the main methods favored by researchers because of their strong robustness and generalization ability.Although many scholars have proposed relevant detection solutions based on computer vision methods,the complexity of the change detection task itself still makes the field face the following problems:(1)imaging interference of multi-temporal images,this contradiction is more acute in the high-resolution environment,mainly due to the imaging angle of high-altitude satellites and imaging time(such as affected by sunlight),especially for high-rise buildings,which are very easy for the same object to appear with different characteristics in different temporal;(2)incomplete detection of multi-scale buildings,mainly caused by the appearance and size diversity of modern buildings,deep learning networks are hard to learn the semantic information with strong compatibility,and it is difficult to adapt and accurately detect various scales of building entities;(3)the extreme class imbalance problem in the change detection dataset,which is mainly manifested by the large number of negative samples in the change detection dataset,and these negative samples will cause interference to the convergence of the network;(4)incomplete contextual association and semantic information mining,this problem is mainly limited by the computing power of the graphics card,in most cases,the images need to be cropped or scaled before formal detection to ensure successful training,which makes it difficult to avoid the problem of information discontinuity,secondly,the high-resolution images contain more information and involve multi-temporal factors,which makes it difficult for deep learning models to establish the long-range connections.To solve the above problems,this thesis proposes relevant deep learning architectures for change detection based on the challenges encountered in remote sensing image change detection in the engineering context,improves the overall performance of the network,and integrates the relevant methods into the prototype system.The main research contents are as follows:(1)For the problems of imaging bias,multi-scale entities,and class imbalance in high-resolution remote sensing images,the spatial-temporal based multi-head self-attention for remote sensing image change detection is proposed.First,the multi-head attention module in Transformer model is used separately and applied to the features extracted by the backbone network to strengthen the relationship between the two temporal images and reduce the influence brought by factors such as imaging interference.Then,after being processed by the multi-headed self-attention module,the corresponding features are fed into a Multi-part Feature Learning(MPFL)network,and the information implied in the features is further extracted using convolutional kernels of different sizes,so that the network can adapt to change entities at multiple scales,while the impact of change detection data class imbalance on the model is also mitigated by improving the loss function.Finally,the output is performed by the metric module.This method is experimented on two public change detection datasets,LEVIR-CD and DSIFN-CD,and the experimental results show that the scheme has significantly improved the model performance.(2)For the problem of incomplete contextual association and information mining of high-resolution remote sensing images,the Transformer-based change detection method with spatial-temporal information aggregation is proposed.This method uses a two-branch improved Transformer module with shared parameters as the encoder module for extracting two-temporal images,followed by extracting and superimposing multiple layers of features at different scales and differentially fusing the features at each layer before inputting to the decoder to mine spatial-temporal semantic information.Finally,in the decoder,a lightweight two-stage up-sampling convolution is used to generate the final change map.In addition,this method is optimized for the large overhead of the Transformer-based detection method to achieve better results with single GPU.In this thesis,experiments are conducted on two public change detection datasets,LEVIR-CD and DSIFN-CD,and the experimental results demonstrate that the scheme has improvement for the performance of the model with limited overhead.