Change Detection In High-resolution Remote Sensing Images Based On Attention Mechanism And Twin Structur

Remote sensing image change detection has great application value in many fields,such as land cover change,urban development research,environmental monitoring,and disaster assessment.In recent years,with the development of relevant theories and technologies in the field of artificial intelligence,the methods related to deep learning have been gradually applied to change detection.The deep learning method has strong feature extraction and expression capabilities,and can fully mine the deep feature information of remote sensing images.Compared with the traditional change detection methods that are limited by artificial design features,the deep learning algorithm can better express the complex surface conditions in the image and get better change detection results.However,with the development of highresolution images,the current deep learning methods have some shortcomings when faced with rich spatial features,diversified scale features of ground objects and a large amount of remote sensing image data.This paper analyzes the shortcomings of the existing deep learning model methods,studies the network model optimization methods,and constructs two kinds of change detection models based on Siamese networks and attention mechanism from the multi-scale features of ground objects.The main research conclusion and contents of this paper are as follows:(1)High-resolution remote sensing image dataset production.This paper makes use of the remote sensing image data of Gaofen 2 in Kunming and Qujing in central Yunnan to produce the dataset.The main contents include data collection,preprocessing,image fusion,registration,sample clipping and change information labeling.In the process of change information labeling,only the disappearance and appearance of objects are regarded as changes,ignoring the changes caused by seasonal differences,brightness and other factors.(2)A Multi-Scale Attention Mechanism Based Change Detection Model.To address the issues of missed detection of different scale objects and incomplete region detection in remote sensing image change detection,a high-resolution remote sensing image change detection model based on Siamese network and multi-scale attention mechanism(Multi-scale Attention Siamese Network,MASNet)is proposed.The model utilizes a Siamese Res Net-50 network to extract features from different time images,and then applies attention modules to feature maps of different scales to generate multi-scale feature representations.By improving the contrastive loss function,the learning of change features is enhanced,and the problem of class imbalance between invariant and change samples is improved,thereby improving the accuracy of change detection.The model achieved good results on a self-made dataset.Compared with six publicly available models in the comparative experiments,MASNet improves the F1 score and recall rate R on the self-made dataset by at least 2.92% and 4.52%,respectively.(3)A Multi-Scale Feature Transformer-Based Change Detection Model.To address the issues of insufficient semantic information extraction in deep learning networks,loss of highorder multi-scale feature details,and non-prominent image difference information,we propose a high-resolution remote sensing image change detection model based on the Siamese structure and multi-scale feature Transformer(Multi-scale Feature Transformer Siamese Network,MFTSNet)on the basis of the MASNet model framework.We replace the Res Net-50 with Res Net-18 in the feature extraction module to reduce the number of parameters in the feature extraction stage while ensuring the feature extraction effect.We also replace the positionchannel attention in the attention module with Transformer-based attention and design a semantic Transformer module to capture semantic information from feature maps of different levels.The model introduces Grounding Transformer and Rendering Transformer modules to enhance the acquisition of low-level and high-level semantic information,supplement highorder multi-scale feature details,and global context relationships between different spatial positions and channels,further improving the accuracy of change detection and optimizing the completeness of detected objects,region internals,and edge details.In the experimental analysis stage,we compare MFTSNet with MASNet and other advanced change detection models on four public datasets,and use a large number of comparative experiments,ablation experiments,and parameter analyses to verify the effectiveness of MFTSNet.The experimental results show that,compared with seven other public models in the comparative experiment,MFTSNet improved F1 score and Io U on the four public datasets by at least 2.55%,0.113%,0.369%,4.216%,and 3.164%,0.188%,0.494%,and 6.202%,respectively.