Research On Classification Of GF-2 Remote Sensing Image Based On Improved UNet Network

High-resolution remote sensing images contain more information,and traditional pixel-based classification methods are not suitable for high-resolution remote sensing images.At present,the object-oriented classification method is the main method for extracting high-resolution remote sensing image information.It can comprehensively utilize various features such as the geometry,texture,and spectrum of each object to extract the required information.However,at present,the object-oriented classification method is not mature enough,and it still needs more human participation in the optimal segmentation parameter acquisition and feature optimization.Therefore,it is still necessary to explore an intelligent and automated image classification method to be fast and efficient extraction of land use information.In recent years,deep learning has played an important role in visual analysis,language and image recognition.Because of its more internal hierarchical structure,it can analyze and reason from the known information of the image to more abstract information.In this paper,deep learning is applied to high-resolution remote sensing image classification,and the UNet model is selected.This model was originally used for medical image segmentation and solved the problem of binary classification.There are still difficulties in extracting the information of complex remote sensing image cells and model training.Problems such as slow speed and difficulty in convergence.For the above problems,this paper studies as follows:(1)Summarize the development status of traditional methods and deep learning methods in high-resolution remote sensing image classification,analyze and sort out the advantages and disadvantages of deep neural network in multi-classification,and then propose a high-resolution remote sensing image classification method based on deep neural network.(2)Understand the general situation of the study area,formulate the classification system of the area,collect images,vectors and field measurement data.According to the GF-2 image data,preprocessing work is carried out.Comparing multiple image fusion methods,it is proved that NNDiffuse fusion achieves better results in this study area in terms of spatial resolution and retention of spectral information.(3)Explain the basic structure layer of the convolutional neural network,and select the classification model UNet network of this article according to the actual situation of this area.Analyze the shortcomings of the basic UNet model,and on this basis,improve it to a deeper model,so that it can extract deeper features of more complex images,and change its convolution method and join the Dropout layer,greatly reducing model parameters and improving Training and testing speed to avoid overfitting.(4)Conduct experiments based on the improved UNet model,evaluate and analyze the experimental results,and compare with the other two traditional classification methods.Experiments show that for the classification of larger areas,object-oriented classification is affected by various factors such as segmentation results,feature selection,and classification algorithms,and the classification effect is not easy to achieve good results.And deep learning has the ability to extract deep features,and does not require too much human participation.By comparative analysis,the overall accuracy obtained by this method is the highest.Except for grassland and water,the accuracy of each feature type is also higher than that of the other two methods.Moreover,the boundary of the classification map obtained by the UNet classification method is closer to the real situation of the ground features,there are few broken patches,and it is smoother and more beautiful.