Research On Semantic Segmentation Of Remote Sensing Image Based On Deep Learning

In recent years,many countries have gradually increased the number of satellite launches in order to seize the "air control rights".Remote sensing is an important technical means to acquire "air control right".By means of sensors mounted on satellites,the imaging of long-range targets is realized.In order to reflect the value of remote sensing image,it is necessary to extract the key information.Semantic segmentation of remote sensing image is an important prerequisite for information extraction of remote sensing image.At present,the research on semantic segmentation of remote sensing images generally tends to be combined with deep learning techniques.Although these studies have improved the segmentation effect compared with the traditional algorithms,most of the research is not mature,and the segmentation accuracy and running time cannot be deployed on a large scale and in real-time environment.Based on deep learning technology,this paper will continue to explore the semantic segmentation of remote sensing images based on deep learning.The main contents of this paper include:(1)Considering that the training data in the semantic segmentation of remote sensing images based on deep learning are all manually annotated and there is no public data set with annotation as large as ImageNet for research use,this paper proposes to use the loss sensitive conditional generation antagonistic network as the architecture,and construct the UED generator by using the structure and skip connection of the U-Net network encoder and decoder for referrence,so as to construct the LS-CGANs algorithm for remote sensing image enhancement.(2)Aiming at the problems of unbalanced distribution of segmented ground objects in remote sensing images,easy overlapping of edges of different segmented objects,small scale of individual segmented objects and difficulty in distinguishing texture details,the U-Net network model which performs well in biological cell segmentation needs to be improved.This paper proposes an improved U-Net algorithm by consturcting Meu and Mdu through some optimization strategies such as batch normalization,deconvolution and skip connection,and using SeLU as the activation function instead.This algorithm improves the segmentation accuracy on the premise of small sample training.(3)Image semantic segmentation is pixel-level.Faced with too fine label remote sensing image data,the trained model is fragile,and the generalization ability is poor.In this case,U-Net and SegNet are mixed,and the DU-SegNet algorithm is proposed.The convolution operation is changed to hole convolution,which retains more details of the original image,further improving the segmentation accuracy of the model.Experiments show that the proposed LS-CGANs algorithm can effectively enhance remote sensing image data.The improved U-Net algorithm has higher segmentation accuracy on small sample datasets.The proposed DU-SegNet algorithm guarantees the training advantage of small samples,and the segmentation accuracy rate is 92.45%,which has strong practical significance.