Research On Change Detection Of High-resolution Remote Sensing Images Based On Transforme

Remote sensing image change detection refers to that used remote sensing data which is registered in different time of the same surface observation area,combined with the characteristics of detection entities and the remote sensing imaging mechanism,and uses image processing technology to analyze the location,range,category and other changes of the objects in the detection area.Over the years,change detection has been widely concerned by researchers and has been one of the important research topics in remote sensing.The traditional change detection method requires more manual intervention,and the detection effect is poor,which cannot cope with remote sensing big data.How to detect changes accurately,efficiently and automatically has become the focus of current research.In recent years,relied on the powerful feature expression ability and nonlinear modeling ability,deep learning technology provided new ideas for professionals who engaged in researching change detection.This paper uses deep learning technology to carry out research on change detection methods.The specific contents are as follows:(1)Since the convolution operation cannot establish the long-distance dependence in feature space,resulting in the receptive field of convolution neural network is limited.Therefore,this paper proposed the MSTrans Net,which based on multi-scale Transformer network.The multi-head self-attention mechanism has a global receptive field,which can establish the longdistance dependence in the feature space.In order to effectively consider different scale detection entities and irregular geometric structure,introduced the multi-scale module to compensate for the missing multi-scale information in the network.In order to enhance the ability to express local information,add the depth-wise separable convolution to feed forward neural network.The experimental results show that the MSTrans Net is better than the contrast model in quantitative and qualitative performance,with the quantitative results of LEVIR-CD dataset reached 92.13%(Precision),89.72%(Recall),90.91%(F1-score),83.34%(Io U)and99.09%(OA),the quantitative results of CDD dataset reached 96.71%(Precision),96.27%(Recall),96.49%(F1-score),93.21%(Io U)and 99.17%(OA).Moreover,the effectiveness of each module was verified through ablation experiments.However,the stacking of a large number of multi-scale transformer module results in low computational efficiency in MSTrans Net.(2)When maintaining the high accuracy of change detection,improving the efficiency of the network,this paper combined with the CNN and Transformer structure,proposed the PTTrans Net which Transformer structure based on pyramid semantic token.PT-Trans Net uses Res Net18 without maxpool layer to extract hierarchical features to retain fine-grained information such as edges in high-resolution features.In addition,PT-Trans Net regards features as some semantic token to shorten the sequence length and improve the operational efficiency of Transformer structure.In order to make semantic token have multi-scale representation,used spatial pyramid pooling to generate multi-scale features,then flattened and concatenated the matrix to generate pyramid token,before that,introduced Convolution Block Attention Module and deep supervision strategy to enhance the token’s feature representation ability.In order to enable semantic token to learn multi-level semantic information and enhance anti-interference ability against easily confused information,concatenating the four levels of semantic token to form the super-sequence,and interact multi-level semantic information in the Transformer encoder.Used the semantic token with rich information to calculate the long-distance dependence of original feature space.Finally,integrating multi-level features through layer-bylayer up-sampling and element-by-element addition,and generated the final change detection results.The experimental results show that PT-Trans Net performed better than the contrast model in the public dataset,and the relevant network’s structure effectively improved the detection effect of PT-Trans Net.The quantitative results of LEVIR-CD dataset reached 92.48%(Precision),91.06%(Recall),91.76%(F1-score),84.78%(Io U)and 99.17%(OA),and the quantitative results of WHU-CD dataset reached 94.38%(Precision),93.05%(Recall),93.71%(F1-score),88.17%(Io U)and 99.52%(OA).Through analyzing the accuracy,parameter quantity and computational efficiency of all models,it can be seen that PT-Trans Net achieved the balance between accuracy and operation efficiency.