Research On Cloud Detection Algorithm Of Remote Sensing Image Based On Multi-level Encoding And Decoding Structur

Cloud detection in remote sensing images is an important task because cloud coverage may interfere with the extraction and analysis of surface objects.The cloud detection algorithm can automatically identify the cloud in the image and separate it from the image,making the follow-up analysis tasks more accurate.In addition,the cloud detection algorithm can also provide help for the fields of earth resource research and natural disaster monitoring.At present,field of cloud detection algorithm is still facing a series of problems.For example,the boundary of the cloud layer is often an irregular geometric figure.This information is very easy to be ignored during the cloud detection process.In addition,the location of the cloud layer often has complex and diverse ground object information,and the distribution of the cloud layer is irregular and discontinuous.However,the existing segmentation network has defects in the extraction of global information,and it is easy to lose the relative position information between the cloud and the cloud shadow.In order to solve the above problems,the end-to-end training mode of convolutional neural network is used to solve the cloud and cloud shadow segmentation problem of different remote sensing satellites in the background of complex ground objects.On the one hand,this paper proposes a convolutional neural network based on U-shaped codec structure for remote sensing image cloud detection.The network makes full use of the semantic information existing in remote sensing images and uses the high-frequency feature extraction module,multi-scale convolution module,spatial prior attention module and spatial channel attention module to greatly improve the detection accuracy.On the other hand,considering the actual engineering needs,so that the model can balance the reasoning speed and detection accuracy in the cloud detection task,this paper proposes a more lightweight neural network.The network also exhibits good segmentation performance on multiple datasets and has potential application value.