Method Research Of High Resolution Remote Sensing Imagery Classification Based On U-Net Model Of Deep Learning

High resolution remote sensing images record the detailed spatial layout,geometric structure,texture,and other characteristics of ground objects.They are widely used in many application fields including remote sensing mapping,precision agriculture and urban planning.Although the high resolution remote sensing images provide high-quality information,the characteristics of high intra-class variance and relatively lack of spectral information also pose new challenges for the efficient and accurate classification and interpretation of remote sensing images.On the one hand,high resolution images show highly detailed information,increasing the intra-class variance of objects while decreasing the inter-class variance.On the other hand,the bands of high resolution remote sensing images are limited and the richness of spectral information is insufficient,which increase the difficulties to classify remote sensing images.Therefore,how to achieve efficient and accurate classification of high resolution remote sensing images is an important issue that needs to be solved urgently.Recently,in traditional high resolution remote sensing imagery classification methods,the most commonly method is based on object-oriented classification theory and machine learning algorithms.However,it needs to manually participate in segmentation parameter selection and object feature selection,which is timeconsuming and labor-intensive.What's more,the shallow structure model in machine learning can not obtain better classification results.Deep learning is an emerging technology in the field of image recognition in recent years.It can automatically learn the deep features of images to make accurate classification decisions,bringing new opportunities for better high resolution remote sensing image classification results.This paper studies the method of U-Net model for high-resolution remote sensing image classification in deep learning,but the current researches does not fully consider the impact of multi-source data on the classification accuracy,which it is easy to cause low classification accuracy due to the lack of spectral information.At the same time,there is a problem of subtle misclassification of the classified images and smooth boundary of the objects.Above all,the following works has been carried out in this paper:(1)Focusing on the topic of classification of high-resolution remote sensing imagery,the relevant research progress at home and abroad are reviewed the system reviewed systematically.Analyzes and summarizes the current domestic and foreign high-resolution remote sensing image classification methods,and explains the advantages of using deep learning to achieve accurate and efficient classification of remote sensing images;(2)Aimed at the problem that the high resolution remote sensing images has few bands and the feature richness the model learned are limited,the combination of multisource data of DSM,NDVI,nDSM with original images are studied and participate in the image classification process of the deep learning U-Net model,then find the most effective multi-source data combination for improving classification accuracy;(3)Aimed at the problem of subtle misclassification of the classified images and smooth boundaries of the objects,a method for image post-processing using fully connected CRFs is introduced.The optimal parameter for fully connected CRFs is obtained using the grid search method to achieve optimal post-processing effect;(4)A comparative analysis of the results of this method and traditional objectoriented machine learning classification method as well as fully convolutional network in high-resolution remote sensing image classification experiments was conducted to verify the effectiveness of the proposed method.The research shows that: nDSM image data has the most significant impact on the U-Net model classification accuracy,and the included object elevation information can significantly improve the classification accuracy of impervious surfaces,building,low vegetation;fully connected CRFs can eliminate subtle misclassification phenomenon and get a more detailed objects boundaries;the results of this paper have certain reference value for improving the classification accuracy of high-resolution remote sensing images.