Cloud Detection Of Multi-angles Remote Sensing Image Based On Deep Learning

Remote sensing satellite images have been widely used in fields such as land surveying,positioning and navigation,environmental detection,disaster prevention,and ocean utilization.However,the presence of clouds will not only affect the ground scene,but also change the spectrum and texture information on the image to a certain extent,so the cloud detection method is also one of the most popular research directions in remote sensing image processing.In recent years,convolutional neural network has made rapid developments in image processing,which also brings new inspirations to remote sensing image cloud detection research.This thesis mainly studies how to improve the cloud detection accuracy in remote sensing satellite images with multi-angle polarization information.Specifically,this thesis takes multi-angle screening and target detection of multi-angle polarized remote sensing image as the research content.The multi-angle screening of remote sensing image can obtain more accurate target category label,and the target detection can obtain the target pixel.This research has certain theoretical significance and practical valuesIn the cloud detection of multi-angle polarization remote sensing satellite image,it is difficult to select the Angle information to be used in the experiment.Aiming at this problem,this thesis uses the SegNet network to study the influence of different Angle information on the cloud detection accuracy and the advantages of different Angle information fusion.In the target detection,aiming at the complicated problem of manually selecting multi-angle information,this thesis designs a new remote sensing image cloud detection method based on SLS(Senet-LiteSegnet)network,and the detailed work is as follows(1)In view of the shortcomings of the cloud detection using single-angle remote sensing satellite images in the past,this article is based on the advantage of multi-angle remote sensing satellite image,and uses the information from multiple perspectives in the images to better express the semantic information of the image.Firstly,different angles of the information of remote sensing satellite images are sent into the network for learning,in order to obtain different Angle precision training network,and we discuss the effect of single Angle information for cloud detection model,Then,according to the testing results that will be different Angle information fusion,we design a SegNet network based on remote sensing information from multiple perspectives,this method will be different Angle information fusion.The training of multiple multi-angle cloud detection models proves that the cloud detection model trained with multi-angle information has better detection accuracy than the single-angle cloud detection model.On the self-built dataset POLDER3,the multi-angle information cloud detection model proposed in this thesis achieves a global detection accuracy of 91.39%,which is 1.53%higher than the single-angle information cloud detection model(2)Aiming at the problem of selecting training models with different angles that will produce great differences,this thesis designs a new remote sensing satellite image cloud detection model based on the SegNet network structure of semantic segmentation model and the advantages of SEnet,the channel attention mechanism.Taking SEnet as SegNet network layer weight training,the different aspects of using the layer weights "channel" given by different point of view,the design of network has better prediction ability,test results more close to the real sample label,at the same time due to the weight of training can better express the sample information,reducing the previous single Angle remote sensing satellite image selection of differences in the different point of view.Compared with previous proposed multi-angle remote sensing image based on SegNet cloud detection model,the improved SSE network has better effect of semantic expressionMultiple comparative experiments on self-built dataset POLDER3 show that the method proposed in this thesis is able to obtain better results of multi-angle remote sensing satellite image cloud detection,thereby verifying the effectiveness of multi-angle information and SLS network.