A Study Of Remote Sensing Image Ground Segmentation On Deep Learning

Remote sensing image segmentation plays an important role in the field of satellite image research,which is the important foundation of ground object segmentation.Remote sensing image is different from the ordinary three-channel image,it has rich channel types and a large number of spectral characteristics,which provides a reliable theoretical basis for marine pollution monitoring and crop yield estimation.However,the disadvantages of remote sensing image segmentation are gradually emerging.Common pixel classification methods,such as support vector machine and random forest,only use spectral variable features,rarely involve the spatial information of the image.This leads to the lack of the ability to abstract the features of remote sensing images.When dealing with the non-linear relationship with high spatial complexity,the accuracy of ground object segmentation is often low.Therefore,this topic will use the method of deep learning to extract the features of remote sensing image,and combine the deep neural network with the attention mechanism,as well as the method of integrating multiple deep neural networks to segment remote sensing image,which improves the accuracy of ground object segmentation.The main research work is as follows:1.A method of terrain segmentation based on UNET integrated network is proposed.Firstly,10 UNET models are trained with the sample data of 10 kinds of ground objects,and then several trained UNET network models are integrated with the weighted average method to get the UNET integrated network model.The UNET integrated network model reduces the variance of neural network,and has higher decision-making ability and lower error rate than single model.Finally,compared with UNET,FCN and segnet improved by 17.0%,19.0% and 9.0% respectively.2.A method of terrain segmentation based on UNET using convolution block attention is proposed.In order to solve the problem of small and hard to segment objects,we first use the methods of flipping,zooming and deformation to enhance the data of small objects.Because the image resolution after zooming and deformation is low,we will use the super-resolution network to improve the resolution of the image and get the high-resolution sample data.Then we use the convolution block attention module CBAM(convolutional block attention module Block attention module)iscombined with UNET to get UNET attention model.Using the attention module of CBAM,we can select the channel and space area of the convolution neural network,improve the attention ability of the convolution neural network,and avoid the influence of the background on the ground object segmentation.The experimental results show that the segmentation accuracy of UNET is 19.0%,21.0% and 8.0%higher than that of full convolution neural network,segnet and UNET respectively,especially in the segmentation of small objects such as vehicles and lakes,which is3.0% higher than those of the three methods.3.Developed the remote sensing image segmentation system,in order to better display the results of remote sensing image segmentation using this method.Through this system,we can more clearly and intuitively see the comparison of different methods in the same remote sensing image segmentation results.