Research On Semantic Segmentation Algorithm Of Remote Sensing Image Based On Convolutional Neural Network

With the rapid development of remote sensing technology,both national defense and commercial applications have an urgent need for automatic extraction of remote sensing image information.Semantic segmentation is an important prerequisite for remote sensing image information extraction.Based on this,more extensive research and applications can be performed.Semantic segmentation is the process of labeling each pixel in an image with semantic labels,and can simultaneously complete the image pixel classification and image object segmentation processes.In recent years,image semantic segmentation mainly uses deep learning algorithms to automatically extract information.Although there are many researches and applications of deep learning in high-resolution remote sensing images,due to the complex scenes of remote sensing images,it poses a challenge to the semantic segmentation algorithm based on deep learning.Based on a systematic analysis of deep learning,this thesis delves into the semantic segmentation algorithms of high-resolution remote sensing images.By improving and optimizing the classic semantic segmentation model,it is better adapted to the semantic segmentation tasks of remote sensing images.The main research contents of the paper are as follows:（1）Exploration of semantic segmentation model.Aiming at the category imbalance problem of remote sensing image semantic segmentation,by modifying the cross entropy loss function,different weights are assigned to the corresponding categories.In order to make full use of the context information,the Vortex Pooling module is used to integrate multiple prediction results using ensemble learning.To eliminate the noise generated in the stitched images,binary morphology is used for post-processing.In addition,for highresolution images with few bands and insufficient spectral information,DSM（Digital Surface Model）data is introduced and inputted into the network for training along with high-resolution remote sensing data to achieve semantic segmentation of multi-source remote sensing data.（2）Exploring the semantic segmentation model of single source remote sensing image based on Deeplab V3+.First,by modifying the cross-entropy loss function,using the Vortex Pooling module instead of the ASPP（Atrous Spatial Pyramid Pooling）module to improve the context information of the image;then,multi-scale input is used to extract features at different scales and feature voting is used to perform feature fusion to improve the accuracy of image segmentation.Finally,the prediction results are post-processed using the morphological mid-open operation.Train on the dataset of the CCF（China Computer Federation）competition and compare it with other classic semantic segmentation algorithms.Use commonly used evaluation indicators（accuracy rate,recall rate,F1 index and average intersection ratio）to evaluate the experimental results,and analyze and compare the segmentation effects of I-Deeplab V3+ algorithm,Deeplab V3+algorithm,Fcn algorithm,U-Net algorithm,and Seg Net algorithm.（3）Exploring the semantic segmentation model of remote sensing image based on U-Net.In view of the lack of high-resolution image bands and the lack of rich spectral information,the research combined DSM data with the original image.First,the original image and DSM data are input to the network for feature extraction,and the extracted features are fused.Then,the Vortex Pooling module is used to improve the context information.Finally,the improved U-Net is used to implement the semantic segmentation of dual-source remote sensing data.（4）Experimental results show that the improved Deeplab V3+ algorithm in this paper makes full use of context information,effectively reduces misclassification,and makes segmentation boundaries more accurate.Especially for roads,rivers and other linear targets,the ability to capture is stronger,the MIo U on the entire test image can reach85.21%,which is significantly better than the FCN algorithm,Seg Net algorithm,and UNet algorithm.The improved Unet algorithm added to the elevation information in DSM can help distinguish low-lying vegetation and trees,vehicles and other easily confused features,further confirming the validity of the multi-source remote sensing data combination.In addition,by increasing the amount of data and the Dropout layer,the phenomenon of overfitting is effectively reduced,and the accuracy of semantic segmentation is further improved.The average F1 value of all categories of the algorithm on the test set reached 83.60%,and the MIOU reached 72.30%.