Research On Building Extraction From Remote Sensing Images Based On Convolution Neural Network

In recent years,with the development of technology,the source of remote sensing image data is constantly updated,and it becomes more and more important to extract building information accurately from remote sensing image.Traditional automatic building extraction methods often rely on artificial design features.However,the effect of traditional methods is often limited due to the diversity of the shape,size,position and color of buildings in remote sensing images,and the high resolution remote sensing images contain a lot of noise information.In recent years,new technologies represented by deep learning methods such as convolutional neural networks have been introduced into remote sensing images,providing a new idea for extracting accurate building information from massive remote sensing images.Deep learning technology can automatically extract features and information from high-resolution remote sensing images,replace traditional artificial feature design methods,and use models to replace complex data processing processes.Therefore,deep learning technology can effectively solve the problem of extracting accurate building information from remote sensing images.This thesis mainly studies and analyzes the convolutional neural network method based on deep learning,which has achieved significant improvements in accuracy compared to traditional methods.However,this method still has many issues that need to be addressed.In order to address some of these issues,this thesis conducted the following related research:1)This paper proposes a novel method of building extraction from remote sensing images to address the issues of insufficient feature representation and unsatisfactory extraction results.Specifically,we adopt U-net architecture and integrate convolutional block attention module to alleviate the impact from background and noise interference.Experimental results on the WHU dataset demonstrate that our proposed method outperforms the baseline U-net model and achieves more accurate building segmentation owing to its superior feature extraction capability.2)Thesis proposes a fused CNN and Transformer encoder-decoder segmentation model to overcome inherent limitations of CNN models and relatively poor ability to model long-range dependencies.The proposed model introduces local enhanced position coding to better process local positional information and employs a crosswindow attention mechanism to achieve powerful modeling capabilities while minimizing computational costs.Comparative experiments conducted on a dataset demonstrate the effectiveness of the proposed model,indicating that the fused encoderdecoder model using Transformer and CNN is capable of accurately extracting buildings from remote sensing images.