| In recent years,the rapid development of optical satellites has become an indispensable part of a stable and efficient earth observation system,which has brought great convenience to human observation of the earth.Optical satellite images have the advantages of wide coverage,low acquisition cost,short repeated observation cycle and large amount of information and data.They are widely used in many fields,such as national defense security,fishery monitoring,environmental protection,natural disaster early warning,resource investigation and so on.However,as a kind of pollutant in satellite images,cloud greatly reduces the quality and utilization efficiency of satellite images.Research by the international satellite cloud climate project shows that clouds cover more than 65% of the world.The existence of cloud will reflect many band signals of optical satellite,which makes it difficult for optical satellite to obtain the complete surface information under the cloud,resulting in the lack of satellite image information and difficult to use.Therefore,cloud detection is a necessary measure to improve the utilization and application range of remote sensing images.The cloud detection of optical satellite image is actually an image segmentation task with cloud as the foreground and other ground objects as the background.The rapid development of deep learning algorithm brings opportunities for efficient and accurate image segmentation.At present,most cloud detection algorithms usually carry out complex model construction and feature extraction based on a large amount of prior knowledge.In the face of complex underlying surface,manual intervention is often required to effectively capture image features in order to achieve the purpose of cloud detection.The deep learning algorithm represented by U-Net network can mine the characteristics of data set more accurately,update the model parameters iteratively by manually labeled samples,and extract the shallow and high-level semantic features contained in the image through different network layers,so as to effectively detect the cloud in the image.Therefore,based on the theory of deep learning and the lightweight U-Net network,this paper studies the problems in cloud detection of optical satellite images.The main research contents and innovations of this paper include the following two aspects:1.Aiming at the problem that most deep learning algorithms need a large number of manually labeled data samples and have poor discrimination effect on complex underlying surfaces such as snow / ice,this paper constructs a U-Net network cloud detection model based on attention mechanism based on lightweight U-Net network.By adding a spatial attention module to the upper sampling layer of the network,this method improves the ability of the network to pay attention to the cloud,so as to improve the accuracy of cloud detection.Firstly,the encoding layer obtains the shallow and deep semantic features of images at different scales through multiple downsampling operations;Then,the attention module is added to the upper sampling layer,and the feature map obtained by down sampling at the same layer of the network is fused after full convolution operation;Finally,the classification results are output to achieve the end-to-end cloud detection target.The experimental results show that the overall accuracy of the model verified on landsat8 CCA data set is 92.91%,which is8.12% higher than that using only U-Net model,MLP（Multilayer Perceptron）model13.48% and RF（Random Forest）model 13.17%,and realizes efficient optical satellite image cloud detection.2.Based on the U-Net network model based on attention mechanism,from the perspective of strengthening the ability to capture edge information,deformable convolution operation is introduced,and a U-Net network cloud detection model integrating attention mechanism and deformable convolution is constructed.The everchanging shape of cloud is an element that can not be ignored in cloud detection,and deformable convolution can adapt to a variety of shapes,which is more effective for learning the edge features of irregular objects.Therefore,this paper uses deformable convolution instead of ordinary convolution in order to improve the attention to the shape of cloud and realize the effective extraction of cloud edge features.The experimental results show that the overall accuracy of the model verified in landsat8 CCA data set is 93.10%,which is 8.31% higher than that of u-net model,MLP model13.67% and RF model 13.36%.The research shows that the two algorithms improved based on U-Net model in this paper can achieve cloud detection targets with small data sets（training data sets less than 300 samples of 256×256 pixels）and complex underlying surface types.The u-NET network based on attention mechanism has the advantages of higher overall accuracy,lower false detection rate and missed detection rate;The U-Net model integrating attention mechanism and deformable convolution has higher overall accuracy,and is more effective in detecting thick clouds,which provides a new perspective for realizing the accuracy of cloud detection. |
