Research On Multi-spectral Remote Sensing Image Classification Based On Improved U-net

Accurate and efficient classification of multispectral remote sensing images has been a hot research topic in the field of remote sensing.Multispectral remote sensing images contain rich spectral information,which can be used to identify and classify features.In the past 30 years,many traditional methods of feature extraction and classification are proposed,such as least square classifier,K-means classifier,support vector machine classifier and principal component analysis based on dimension reduction.However,none of these methods can extract the deep abstraction features of the objects in multispectral remote sensing images.Therefore,deep learning has come into being to solve this problem.Compared with the traditional classification methods,artificial neural network has a strong learning ability,has some fault tolerance,and does not need to assume the probability model and other characteristics,suitable for complex scene features extraction and spatial pattern recognition and other types of problems.In order to solve the problems of low accuracy and low efficiency caused by the traditional classification methods for multi-spectral image classification,the U-net network can effectively solve these problems.U-net can construct deeper network structure,increase training parameters of the model and make the model learn more features of multispectral image data.Setting the corresponding network parameters for specific scenarios can greatly reduce the network Of the calculation time to improve the accuracy of the network,and then improve the classification results of multi-spectral image features.How to use the improved U-net neural network to classify multi-spectral remote sensing image features is the key point of the dissertation.The main research contents are as follows:1)In this paper,we first summarize the traditional multispectral image classification methods,such as supervised classification,unsupervised classification and semi-supervised classification,and analyze the problems such as lack of precision,low efficiency and low fault tolerance of these methods in feature recognition,The important role of convolution neural network in multi-spectral image classification is derived.It is proved that convolution neural network is feasible and necessary for multi-spectral remote sensing image classification.2)Secondly,the concept and improvement of U-net network structure are proposed.The improved U-net network is proved to be suitable for multi-spectral image classification.The multi-spectral image preprocessing is emphasized.Compared with other multi-spectral Image classification method,training data preprocessing can greatly improve the computational speed of the network,so that the network model can learn more data features and improve the recognition rate of the network model.Then it narrates the training process of U-net network,analyzes the model structure,parameters and evaluation index of the network.3)Finally,the importance of postprocessing the experimental results is described emphatically.In many methods that use neural networks to classify,they are only classified by the network model without any post-processing of the classification results,resulting in the accuracy of classification results Poor and poor classification.In order to solve this problem,we use image processing method to further process the experimental results.Experimental results show that the post-processing of experimental results can effectively improve the accuracy of experimental results,providing a new direction for the classification of multi-spectral remote sensing images.