Multi-scale Semantic Segmentation Of Remote Sensing Image Based On Deep Residual Network

The segmentation of Remote sensing image is a necessary process for image interpretation and analysis tasks.With the recent increase in the resolution of remote sensing images and the advantages of deep learning in feature expression,the classification and segmentation of images are no longer relying purely on spectral information,but are further added to the appearance-based image feature to facilitate a transmition from land-use classification to objects-level scene understanding.In this work,image semantic segmentation based on deep network model has become the main research trend of high-resolution automatic semantic segmentation of high-resolution remote sensing images.Semantic segmentation based on full convolutional neural network is widely used in medical imaging,streetscape photography,remote sensing imaging,and other fields.The full-convolutional network structure can infer full-resolution tag data without downsampling,deconvolution,or interpolation.In order to make better use of image features for automatic segmentation of high-resolution remote sensing images,a multi-scale semantic segmentation model based on deep residual network is proposed in this paper,aiming at improving the segmentation of objects with different scales for small sampled remote sensing image data sets.Firstly,we utilize Residual Network(ResNet)and transform it to fully convolution networks(FCN),then random multi-scale data enhancement is performed for small sample data,which improves model classification accuracy and robustness.Limited by the coarser results of the full convolutional network segmentation,atrous convolution is introduced into the up-sampling process to further refine the resolution of the feature map,which improves the accurancy of the output label map.The proposed method is suitable for multi-scale high-resolution remote sensing image labeling problems.In order to study the validity of the multi-scale residual network model applied to the automatic semantic segmentation of remote sensing images,the experiment was based on two standard semantic segmentation data sets of ISPRS 2D Vaihingen and Potsdam,and a series of experimental programs were designed considering different network parameters and improved methods.After analysis,the following conclusions are drawn:(1)On the semantic segmentation datasets of two different remote sensing images,the classification accuracy of this model is competitive.The classification accuracy of the image segmentation results of the Vaihingen dataset is 90.4%,and the classification accuracy of the image segmentation results of the Potsdam dataset is 90.3%.The results show that the model has good segmentation result for remote sensing images of different scenes.(2)Experiments with different model input and training parameters show that:training samples extracted according to different input image block sizes and coverage will affect the model segmentation effect,and the appropriate sample size and sample coverage rate can help improve the segmentation accuracy of regular shape objects.In addition,choosing appropriate learning rate strategies and batch size training parameters during network training can help improve the accuracy of model training.(3)The sample expansion experiment shows that for the urban scene segmentation task of remote sensing image,the multi-scale data Data augmentation method can effectively improve the accuracy of various types of object.Compared with other segmented objects,the vehicle class with smaller sample size and smaller scale can also get higher classification accuracy and segmentation results.(4)The model improved by atrous convolution strategy makes the segmentation result more refined while improving the accuracy of the model.However,when the multi-scale atrous convolution strategy is applied to the segmented objects in this paper,the accuracy of the model is not improved,so for this data set,Only atrous convolution is used for improvement.