Building Change Detection Technology Based On Convention Neural Network In High Resolution Multispectral Remote Sensing Images

The rapid development of China's social economy has accelerated the spatial expansion of the city.In the process of urbanization,buildings as an active urban element,have a large number of updates,so the rapid and accurate building change information extraction has great significance for the business of urban planning.With The rapid development of satellite remote sensing technology,there has a large amount of remote sensing data,and remote sensing images have become an important data source for building change detection technology.High-resolution multi-spectral remote sensing images contain rich feature information,but the situation of "isomorphism and homoplasmism" is becoming more and more serious.This also poses a great challenge to how to use this information reasonably and effectively for building change detection.With the development of deep learning technology,deep learning can be the new idea for building change detection.In this paper,we transform the building change detection problem into image pixel-level classification problem,and refer to the refinement module of optical flow network and the multi-level prediction of FPN network to optimize the full convolutional neural network Unet,then base this design a building change detection shcheme.For the remote sensing images,firstly,this scheme performs preprocessing operations such as orthorectification and image registration on the two remote sensing images.Secondly,the remote sensing images are segmented and labeled with changes in the new building and demolition of the building.After the training data set is produced,the data used to train the optimized Unet model.Finally,we use another the pre-processed remote sensing image for the predicting by the trained model,and the prediction result is post-processed by morphological method.Then we get the final building change results.The results show that,on the remote sensing image dataset shotted by QuickBird,the overall accuracy of the building change information extraction scheme based on the optimized Unet model is 97%.Compared with the building change detection method based on FCN,SegNet and Unet models,the accuracy increases 2%,the F1 value increases by an average of 0.12.Meanwhile,compared with the traditional building change detection scheme,the scheme has higher automation degree,stronger generalization ability and higher practical value.