Research On Cloud Detection Algorithm Based On Global And Local Feature Aggregation

In high-resolution remote sensing images,clouds and their shadows are very strong signals,which interfere with satellite observation.As an important factor in meteorological and climate research,cloud cluster is closely related to global climate phenomena and sudden natural disasters.Therefore,the fast and accurate segmentation of cloud and cloud shadow is of great significance for the application of remote sensing image.At present,for the traditional cloud detection,remote sensing images often have the problems of thin clouds and broken clouds missing detection,cloud clusters are easy to be confused with high reflectivity objects on the ground,resulting in cloud missing detection and inaccurate segmentation of cloud boundary details.To solve the above problems,this paper first designs a global attention feature fusion network to segment cloud and cloud shadow images.The main network of the network adopts the improved residual structure,and constructs multi-scale features in the residual block to enhance the expression ability of model features.On this basis,an improved hole pyramid pooling module is proposed to further extract multi-scale context information.In the decoding stage,global attention upsampling is proposed,which uses high-level context information to guide the recovery of low-level spatial details.Finally,the edge thinning module is proposed to re predict the scoring probability.In practical application,the model can not meet the needs of low-power mobile devices.Aiming at the problems of large amount of model parameters and long reasoning time,this paper proposes a real-time semantic segmentation model based on bilateral aggregation network.The network adopts a two-way feature fusion strategy,one of which is the edge detail branch to extract the spatial detail information of remote sensing image.The other is semantic branching,which is to obtain high-level context information.Then the two-way guided aggregation module is used to fuse the features of the two channels.In addition,in order to improve the segmentation accuracy,a multi-scale loss function is established in the semantic branch.Experimental results show that the proposed model achieves good results in cloud detection.