Research On Semantic Segmentation Algorithm For Remote Sensing Satellite Images

With the rapid development of remote sensing technology,the remote sensing data is getting more valuable gradually.Both from the aspect of military defense and commercial demand,there is an urgent need for the means of remote sensing image information automatic extraction.Semantic segmentation is the technology of image segmentation and detection by machine,which is an important prerequisite for remote sensing image information extraction.Based on this,it will promote more research work and application in this field.At present,there are not many researches on the semantic segmentation of high-resolution multispectral satellite remote sensing images.There are some simple methods using the optical indicator as threshold to segmentation,and others using some simple deep learning models.These methods can achieve a high accuracy in solving a specific target or a specific data set,but do not have good applicability in different data sets of different objects.A variety of machine learning methods and deep learning models are taken for reference,besides some researches of image segmentation in the field of remote sensing.According to the characteristics of remote sensing images,a technical solution is proposed,including data annotation,model training,predicting results.For the characteristics of high precision,wide range and multispectral information of remote sensing images,band synthesis,image fusion and image cut are carried out during data annotation.During the stage of model training,according to the practical situation,some training techniques like data augmentation and fine-tuning can be used to improve the efficiency of training,and the effect of the model.The output results can be optimized by image smoothing during the stage of predicting results.Finally,the experimental results are evaluated Accuracy and AUC,which is commonly used evaluation indicators in this field.The characteristics and effects of SVM,patch-based CNN,FCN,SegNet are analyzed and compared.The usability of each method of the semantic segmentation of remote sensing images is summarized,and some further research is expected.