Research On Semantic Segmentation Algorithm For Satellite Images

With the rapid improvement of satellite technology,satellite images,especially high resolution remote sensing satellite images have been paid great attention by various countries,and have been applied in different fields.Satellite image can extract the relative position and spatial distribution of various natural elements with its rich information and visual image,which provides great space for the development of target semantic segmentation in both civil and military aspects.At the same time,in the wave of artificial intelligence,deep learning has been greatly developed with the ability of computing,which not only brings great changes in the traditional computer vision and robot,but also brings new solutions in such aspects as finance and medical care.Therefore,deep learning is applied to the semantic segmentation of satellite images,which opening up new ideas for military tactics and civilian business planning.In this paper,we use convolution neural network to classify multi resolution satellite images.The main contents of this paper are as follows:Firstly,it summarizes the basic models of deep learning and three characteristics of deep learning,namely,the simplicity,extensibility and mobility of models.The focus is mainly on the conformation,characteristics,research mechanism and development direction of convolution neural network.For the semantic segmentation of satellite images using the learning features of artificial design,the feature design learning is too complicated and the adaptation range is limited.This paper uses convolution neural network to automatically design and extract features.Based on the typical semantic segmentation network,the semantic segmentation network structure of satellite images is designed,which combines the advantages of the existing Convolutional Neural Network(CNN)and the conditional random field(Conditional Random Field,CRF).In view of the small number of images in the satellite image set,and the uneven distribution between classes,this paper adjusts the context semantic environment in the satellite image segmentation network,and combines the rough feature and the fine feature by increasing the jump connection.At the same time,the conditional random field was added to the network output to make the precision more than 16%.In order to improve the network performance,the sample set is preprocessed and added and includ the multispectral image channel synthesis,and the increase of multi remote sensing imaging index.In the view of the difference between the loss function of the convolution neural network in the semantic segmentation process and the traditional classification network,the loss function of the network is improved and the joint loss function is designed.Finally,the semantic segmentation convolution neural network designed for satellite images is tested in the dataset.The segmentation network obtains higher accuracy in a variety of classifications,including buildings,highways,rivers and paths,and the average accuracy is up to 91% in the sample set water category.Experimental results show that the semantic segmentation network designed in this paper has practical value.