Research On Building Change Detection From Remote Sensing Image Based On Multi-task Improved PSPNet

With the continuous advancement of urbanization,land cover types are changing with each new day.Buildings are the most changeable part of land cover types,and as a key component of geographic information database,they play an important role in urban unauthorized construction management,urban expansion analysis and post-disaster assessment.Interpretation based on high resolution remote sensing image is the main way to capture building information.There are many detection methods for building change based on remote sensing image,and the detection method based on deep learning is widely popular at present.Compared with other change detection tasks,buildings as detection targets have high variability in shape,size,color and height,which poses challenges to detection tasks.Insufficient change samples has always been a problem in the field of change detection.In practical applications,the detection span is often short,and the problem of insufficient change samples is more prominent.Based on this,an improved multi-task twin PSPNet network is proposed in this paper.On the basis of the twin PSPNet network,a multi-scale fusion structure is added to fuse the deep and shallow features of images,which can obtain more hierarchical and different feature information.The segmentation tag of buildings brought by multi-task structure can alleviate the problem of insufficient learning of change features in change detection to a certain extent and help solve the problem of fuzzy contour.In this paper,WHU dataset of high resolution aerial image,GZ dataset of high resolution aerial image and Zhuhai dataset of high resolution satellite remote sensing image are made for experiments.The results show that: 1.The multi-scale PSPNet twin network integrates the semantic information of the deep feature layer and the location information of the shallow feature layer in the way of fusing the feature layer of different sizes,which enables the model to extract more hierarchical and different features.The F1 score of WHU data set reaches 86.3%,0.2% higher than that of UNet++ model.Compared with only using the last layer feature map for prediction,multi-scale fusion structure can effectively extract richer image information,thus improving the accuracy of change detection.2.Based on the multi-scale PSPNet twin network,the multi-task structure is further built for improvement.In the case of limited amount of data,the information of original images of two phases that are rarely used by other change detection models is fully mined,and the problem of imbalance between changing samples and non-changing samples in building change detection subject is alleviated from the model itself.Experiments on module analysis on THE WHU dataset show that PPM module,multi-scale structure and multi-task structure all contribute to the improvement of the final model accuracy.Among them,the multi-task structure has the most outstanding contribution to the refinement of the boundary extraction of the changing region.The main role of PPM is to improve the detection ability of small target ground objects;The main function of multi-scale structure is to help the model to improve the identification ability of pseudo changes caused by bare soil,road and other factors.3.In this paper,the multi-scale PSPNet twin network based on multi-task learning is compared with FC-Siam-conc,FC-Siam-diff,STANet,UNet++ and other advanced models on WHU,zhuhai and GZ datasets respectively.The multi-task improved PSPNet network achieved F1 scores of 90.4%,66.1% and 71.2% in the three data sets,respectively,which were the highest F1 values compared with other models.This shows that the proposed model has good generalization performance.