Building Feature Expression And Change Detection From High-resolution Remote Sensing Image With Attention Mechanism

Since the 21st century,China's urbanization development has entered a period of acceleration.As the most active urban element,buildings have undergone a large number of new constructions and reconstructions.Accurate and efficient expression of building features and extraction of change information are important for urban planning,land and resource management,etc.significance.The rapid development of remote sensing technology in the new century provides a large amount of data for the feature expression and change detection of buildings,which poses a huge challenge for how to effectively use the data.At present,the use of neural networks can express the high-level and low-level features of buildings,but it is impossible to determine how the network constructs the features,the feature expression is not comprehensive,and the building feature expression ability is insufficient.In the building change detection of time series data,the front and back time phase images are taken at different times,and the imaging will be affected by changes in lighting,seasons,etc.There is a phenomenon of "same spectrum with different objects,same objects with different spectra",and the network uses small blocks Image training,lack of spatial context information.In response to the first question,this thesis proposes a framework for expressing the depth features of remote sensing image buildings that incorporates attention mechanisms.The framework uses Resnet18 as the basic network and adds 4 attention modules to the residual modules in the Resnet18.Spatial attention Resnet,Channel attention Resnet,Mixing attention Resnet,and Positional attention Resnet.In response to the second question,this thesis proposes the Pyramid Spatiotemporal Attention Residual Network(PSARNet): First,the Resnet is improved,and the multi-layer features are merged;then the feature map is passed through the spatiotemporal attention module.Compared with the network that does not refer to any spatiotemporal dependencies,the spatiotemporal attention module calculates the weight of the relationship between any two pixels at different time positions to generate more differentiated features,and introduces a pyramid structure,Extract multi-scale features,and finally obtain the building change detection results through classification.This thesis selects New Zealand building identification and change detection data released by Wuhan University to conduct research.In terms of feature expression,the effect of building feature expression is evaluated by calculating classification precision,the accuracy rate,recall rate,and Io U of building feature expression using Resnet18 are92.62%,91.38%,and 85.18%.The attention mechanism is integrated based on the Resnet18.The most accurate method is the expression of building features that integrates the position attention mechanism.Its accuracy,recall,and Io U ratio is94.69%,93.72%,and 89.06%.Compared with the Resnet18,the accuracy rate,recall rate,and Io U are increased by 2.07%,2.34%,and 3.88%,respectively.Experiments show that the position attention mechanism can effectively improve the ability of building' feature expression.In terms of change detection,the recall rate,F1 score,and Io U of the pyramid spatiotemporal attention network built based on the position attention mechanism are 94.79%,92.19%,and 85.52%,respectively.Compared with only using the Resnet18,the recall rate is 65.70%,the F1 score is 76.09%,and the Io U is 61.41%,the precision is greatly improved.