Research On High Resolution Urban Remote Sensing Image Classification Based On Convolutional Neural Network

With the continuous development of remote sensing technology,traditional image classification methods do not perform well in high-resolution remote sensing images.With the rise of deep learning,the classification of remote sensing images based on deep learning methods has become one of the hot spots of researchers.In particular,the excellent performance of convolutional neural network in image classification makes it more and more popular in remote sensing image classification.However,the use of convolutional neural network for remote sensing image classification also faces the following problems:On the one hand,increase the convolutional neural network The depth of the network,the stronger the ability of the network to obtain deep features,the higher the classification accuracy of the image,but as the depth increases,the spatial information caused by pooling,especially the small target spatial information,is more serious.On the other hand,with the remote sensing technology the resolution of remote sensing images continues to increase,and high-resolution remote sensing images contain a large amount of information about ground objects,such as texture,shape,and spatial location.However,due to the diversity and complexity of ground objects,high-resolution remote sensing images usually show for larger intra-class differences and smaller inter-class differences,thus giving class task tremendous challenge.In response to the above problems,this article mainly did the following work:(1)Based on the U-Net model,for the problem of poor classification of small targets in highresolution remote sensing images,the spatial pyramid pooling module is used to replace the pooling and upsampling layers of the original U-Net deep network,which effectively mitigates The spatial information caused by pooling,especially the small target spatial information loss problem,and the use of expansion convolution with different expansion rates has the ability to extract information of different scales.The spatial pyramid module effectively extracts feature maps through multiple expansion convolutions of different scales Multi-scale information of different features improves the classification accuracy of the model.The experimental results show that the overall accuracy of the original U-Net model is improved by 1.95% to 86.93%,the F1 score in the small target "car" category is increased by a maximum of 13.81%,and the F1 score is the highest in all categories,indicating pooling The resulting loss of spatial information has been effectively alleviated.(2)Aiming at the large intra-class differences and small inter-class differences in high-resolution remote sensing images,based on the edge detection idea,a parallel attention branch BA-Unet model parallel to the original U-Net network is proposed.The network shares the weights of the coded convolution layer and uses the contour information of the features to improve the classification effect.Aiming at the problem of extreme imbalance between edge pixels and background pixels in the edge attention branch,Focal Loss is used as a loss function in the edge branch to improve the category imbalance and cause the classifier prediction results to be biased towards the background pixel.The experimental results show that its overall accuracy is improved by 2.06% compared to the original U-Net,reaching 87.04%,and the score of F1 in the small target "car" category is increased by 13.39%,which is slightly worse than the model based on the spatial pyramid pooling module.At the same time,by comparing the classification results,it was found that the contour of the ground edge of the object,especially the edge of the building,has been improved.(3)In order to obtain better classification results,this paper uses the spatial pyramid module and the edge attention branch to jointly improve the U-Net model,and obtains the optimal classification results in this paper.The overall accuracy reaches 87.40%,which is relative to the spatial pyramid pooling module.SPP-Unet and BA-Unet,which introduced the branch of marginal attention,increased by 0.47%and 0.36% respectively.