Remote Sensing Image Change Detection Based On Deep Learning

Change detection of remote sensing images refers to the technique of comparing remote sensing images of the same location at different times to identify differences and analyze change information.Currently,Synthetic Aperture Radar(SAR)systems have achieved continuous dynamic observations over a large range of the globe in allweather conditions,and the SAR images generated are immune to the influence of weather and lighting conditions,making them an important data source in the remote sensing field.SAR image change detection technology is now widely used in land planning,disaster monitoring,urban planning,forest protection,geomorphic change,glacier melting detection,military combat guidance,and more.Compared to traditional change detection algorithms,deep learning has gradually become one of the most powerful tools in change detection technology due to its strong data representation and learning capabilities.Therefore,this paper conducts the following research on how to improve the ability of SAR image change detection.(1)Due to the special imaging mechanism of Synthetic Aperture Radar system,SAR images are severely affected by coherent speckle noise.In order to mitigate the impact of noise on change detection results,this paper proposes a SAR image change detection method based on frequency domain analysis and multiple difference operators.The proposed method first uses multiple difference operators to generate high-quality difference images.The multiple difference operator uses a decision fusion method that combines log-ratio operator and mean-ratio operator to convert the multiplicative noise of SAR images into additive noise,and combines neighborhood information to generate high-quality difference images.Secondly,noise usually exists in the high-frequency part of the image,and processing it in the frequency domain is more advantageous in reducing noise interference.Therefore,a frequency domain analysis module is proposed to process the image in the frequency domain and selectively apply gating linear units to the frequency domain coefficients to effectively suppress the impact of noise on change detection results.The proposed algorithm is experimentally validated on several SAR image datasets and achieves good results.(2)This paper proposes a SAR image change detection method based on a Transformer-based joint network.Due to the insensitivity of frequency domain to local features,a spatio-temporal domain network is introduced to jointly extract image features with the frequency domain network,better addressing the speckle noise problem in SAR images.The center of the image often contains more information,and the spatio-temporal domain network designs a multi-region feature structure to extract central domain features from multiple perspectives,incorporating residual structures to effectively improve the model’s generalization ability.In addition,to better extract key image features and improve the network’s performance,a Transformer based on the self-attention mechanism is introduced.It has strong representation ability,can model the long-range dependency of features,and better captures contextual information to extract global features of the image,thereby improving the network’s performance and making the change detection results more robust.The proposed algorithm was tested on multiple SAR image datasets,and the detection results have demonstrated the effectiveness of the approach.