Research On Himawari-8 Satellite Remote Sensing Image Cloud Detection Technology

Remote sensing images are widely used in areas such as geographic mapping and resource monitoring.However,due to the influence of weather,it is difficult to obtain cloudless images from it.They are susceptible to cloud occlusion during data collection and processing,resulting in reduced quality and utilization of data.In order to remove clouds and improve the effectiveness of it,many scholars have proposed effective methods to detect clouds.Cloud detection of satellite images is very important for quantitative remote sensing research and applications.Therefore,as the first step of cloud removal,accurate cloud detection is conducive to subsequent image processing such as classification,segmentation and change detection,and is also one of the basic work of remote sensing image data restoration.Relying on sensors and a priori knowledge,the simple threshold method is susceptible to some influences such as high reflectivity ground and thus misjudge the cloud.In the convolution and pooling operations,the spatial details of cloud are easily lost in the traditional CNN method,which leads to a decrease in cloud detection accuracy.In view of the above problems,combined with the characteristics of Himawari-8,a semantic segmentation cloud detection technology based on FPN extended network model is designed in this paper,and the three bands with the same spatial resolution are used for corresponding research.In our network framework,hollow convolutions are performed on different levels of feature maps,the details of clouds are obtained by increasing the receptive field,and a feature pyramid is constructed to extract multi-scale and advanced spatial features of cloud images.Then different types of clouds are upsampled and fused,and pixel-by-pixel output classification is used to obtain cloud detection results.Through the qualitative and quantitative evaluation of the results in the three methods,and multiple different remote sensing data were test.It is finally shown that the algorithm has better performance and the accuracy evaluation indexes m IOU and OA are increased by 9.06% and 7.75%,respectively.The visualization results are close to the real cloud image,which is superior to the traditional segmentation method and meets the actual needs.At the same time,it can replace the traditional threshold cloud detection method to detect a large number of samples without relying on artificial experience.This method has the characteristics of low cost and reliable accuracy,and provides a broad application prospect for remote sensing data.