Research On Land Cover Classification Based On Semantic Segmentation

After the earth observation satellite was put into use,the technology of land cover classification by high-resolution remote sensing image began to appear gradually.After decades of development,the land cover classification method of Landsat Image has been improved day by day.With the development of remote sensing technology in recent years,high-resolution remote sensing image has been put into the application of land cover classification.Nowadays,the land cover classification methods for high-resolution remote sensing images are faced with the following problems: on the one hand,the traditional land cover classification methods can't deal with the rich detail information presented in high-resolution remote sensing images.On the other hand,remote sensing images in different regions will have different imaging conditions,which causes the distortion of remote sensing images and other problems.The existing methods are difficult to raise a good solution for generalization.In view of the above problems,this paper proposes the use of semantic segmentation technology to realize land cover classification.The solution of this paper is to use deep learning network for semantic segmentation,and the details of this solution are as follows:First,in order to alleviate the problem of class inconsistency in semantic segmentation,this paper proposes a normalized segmentation network.The gating fusion module not only integrates the shallow spatial features and deep semantic features,but also filters the invalid information and contradictory information in the multi-layer features.After that,the channel attention module further calibrated the fused features.Second,in a broad sense,semantic segmentation can be regarded as an algorithm to determine the boundaries of different semantic entities.The intra-class confusion in semantic segmentation task is caused by the lack of edge information.In this paper,edge information network is proposed to alleviate intra-class confusion.Edge information network uses the modified lightweight residual module to fuse edge features.The loss function of edge information network consists of the up-sampling loss and the auxiliary loss of edge feature output,which are weighted added.