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

Segmentation of remote sensing images is an important task in the interpretation of remote sensing images.With the continuous development of computer technology,the interpretation of remote sensing images using deep learning methods has become mature.The existing neural networks will ignore the spatial relationships between target features in remote sensing images,and the semantic segmentation models combined with convolutional neural networks(CNN)and graph convolutional neural networks(GCN)will cause the absence of feature boundary characteristics,resulting in unsatisfactory boundary segmentation of various types of target features,and the neglect of elevation information The neglect of elevation information will affect the model training effect and semantic segmentation accuracy in areas with large height difference.The remote sensing images generally have large size and contain rich spatial semantic information,and the images exhibit large intra-class variance and small inter-class variance,causing the classification imbalance problem,which leads to the unsatisfactory segmentation effect of small target features in high-resolution remote sensing images.To address the above problems,this paper takes Xining City of Qinghai Province as the study area and adopts two semantic segmentation methods to semantically segment the ground cover types in the study area under different terrain conditions.The main research contents and conclusions of this paper are as follows:(1)In the study area,Huanzhong District,Huangyuan County and Datong Hui Tu Autonomous County have high altitude and large difference in elevation,a semantic segmentation model DGCN combining CNN and GCN is proposed.The model inputs elevation data as parameters into the feature extraction network,which minimizes the problem of too much difference in feature information of the same feature in different elevation environments.GCN learns the target features and their spatial relationship features in remote sensing images.The GCN learns the target features and their spatial relationship features in the remote sensing images,designs the corresponding boundary loss function to learn the effect at the feature boundaries,and finally uses the layered fusion method to fuse and classify multiple feature images layer by layer.The experimental results on ISPRS Vaihingen,ISPRS Potsdam and Deep Globe public datasets show that adding elevation information in the feature extraction stage can improve the model’s ability to acquire feature information of surface features.In the experiments on Huanzhong District,Huangyuan County and Datong Hui Tu Autonomous County of Xining City,the segmentation accuracy of DGCN is significantly higher than other comparison models.It proves that when semantic segmentation of surface cover types in areas with high altitude and large difference in elevation is performed,adding elevation information can effectively improve the segmentation accuracy.(2)A new remote sensing image semantic segmentation network CAS-Net with coordinate attention and SPD-Conv is proposed to address the situation that there are more small target features in the north,west,east and central areas of the study area.The coordinate attention mechanism is introduced into the multiscale module to improve the identifiability of classified objects and the accuracy of target localization in remote sensing images,and the Dice coefficient is introduced into the cross-entropy loss function to solve the accuracy problem caused by the imbalance of data set classification.The experimental results on ISPRS Vaihingen and ISPRS Potsdam public datasets show that in the segmentation of small target features,replacing the stepwise convolution in the backbone network with SPD-Conv and adding the coordinate attention mechanism in the multiscale feature extraction module can reduce the neglect of detail information,which can effectively improve the segmentation accuracy of small target features.The segmentation results of Chengbei,Chengxi,Chengdong and Chengzhong districts of Xining city show that the overall accuracy and average crossmerge ratio of CAS-Net for segmentation of small target features are significantly higher than those of other classical semantic segmentation models.