Research Of Semantic Segmentation Of Remote Sensing Image Based On Improved Unet

Remote sensing image semantic segmentation is an important basic link for interpreting remote sensing image information.In the segmentation and segmentation process,the determination of the number of categories is a key issue and plays a vital role in the further processing of the image.This article focuses on analyzing the multi-class semantic segmentation of high-resolution remote sensing images through the U-Net network model in deep learning theory.Based on 14 high-resolution remote sensing images and their corresponding real-value information,this paper proposes an improved method for U-Net network and compares the advantages and disadvantages of the improvement through experiments.The main work done in this article is as follows:(1)Annotate the remote sensing map data set.In order to train the network model with a probability distribution,One-hot encoding is introduced.In order to solve the problem of insufficient number of samples in the data set,random cropping and rotation are used to increase the data set in image preprocessing.And normalize the acquired remote sensing image.Finally,fill the bottom and left side of the image,so that the training image can enter the network with the same size for training.(2)The Adam optimization algorithm is used for model training on the traditional U-Net network,and an adaptive function LearningRateScheduler is used in the selection of the learning rate to realize the adaptive adjustment of the learning rate according to epoch.Use traditional U-Net network to realize the semantic segmentation of remote sensing images.Afterwards,the improvement function Elu is used to replace the defect that the ReLU function is easy to inactivate the neuron,and the confusion matrix is recorded and the overall accuracy and Kappa coefficient are calculated.The Kappa coefficient is increased by 1.3%under the condition of using Elu as the activation function.Accuracy increased by 1.1%.(3)Propose an improved method for U-Net network,define the currently popular activation function GELU in the function layer by changing the activation function,deepen the network depth to make the deepest point of the network reach 2048,form D-UNet and introduce expanded convolution and dropout The layers and so on carry out network training and semantic segmentation respectively.Through experiments,it is concluded that various improved algorithms have a better improvement for U-Net(ReLU)and U-Net(Elu)networks.The Kappa coefficient has increased by 3.6%,4.2%,and 4.1%respectively compared to the U-Net(ReLU)network;and 2.3%,2.9%,and 2.8%respectively compared to the U-Net(Elu)network.Finally,the three improved methods are added to the U-Net network as a whole.The experimental results show that compared with U-Net(ReLU),the network Kappa coefficient is increased by 5%,and the overall accuracy(OA)is increased by 3.7%;compared with U-Net(Elu)network,the Kappa coefficient is increased by 3.7%,overall The accuracy(OA)is increased by 2.6%.Through the experimental data,it is concluded that the three improved methods proposed in this paper can not only complete the multi-classification semantic segmentation task of remote sensing images,but also improve the accuracy rate compared with the traditional U-Net network model,and can better complete the small data set.Multi-class semantic segmentation task of remote sensing images.