Construction And Application Of Cloud Detection Sample Database Based On Landsat-8 And Semi-Supervised Learning

Cloud detection of satellite images is an important preprocessing step for remote sensing applications.With the development of computer theory and technology,machine learning and deep learning algorithms have been applied to the field of cloud detection and greatly improved the cloud detection accuracy of remote sensing images.In the current researches,how to improve cloud detection accuracy,performance stability as well as computing efficiency on a global basis is still a challenging issue.As for model training,selecting sufficient and representative training samples plays a key role.Most of the training samples are selected by manual visual interpretation,which consumes a lot of resources.Previous studies mainly focused on refining algorithms to improve the accuracy of cloud detection models,and few studies have explored how to automatically extract effective training samples.Moreover,the current machine learning cloud detection models are not universal enough to be applied to other sensors at the same time.To obtain more training samples for machine learning models,this study proposed an erosion and disagreement-based semi-supervised learning(EDSL)model,which is combined with Fmask 4.0 algorithm to automatically extract training samples from Landsat-8 remote sensing images and build a cloud detection sample database.After finishing the accuracy verification of EDSL model,to test the effectiveness of the cloud detection sample database,three categories of training samples including samples from Fmask 4.0 algorithm,samples from the cloud detection sample database and samples from manual visual interpretation are utilized to fit different random forest models(Fmask?RF,EDSL?RF and Truth?RF)for cloud detection of remote sensing images.The result demonstrates that the accuracy of the cloud detection sample database after purification of EDSL model reached 0.991.After obtaining the initial cloud mask of images based on Fmask 4.0,the EDSL model could remove the misclassified pixels in the edge of different cloud objects and the blocky misclassified objects in the initial cloud mask.In the task of cloud detection of remote sensing images,the overall accuracy of Fmask?RF,EDSL?RF and Truth?RF are 0.694,0.909 and 0.913,respectively,which indicates that in the actual cloud detection scenarios,the cloud detection sample database can provide effective training samples for the models.Furthermore,this paper made the comparison between EDSL?RF and Fmask 4.0 in the task of cloud detection.The results show that the performance of EDSL?RF in cloud detection tasks is more stable.To enhance the portability of cloud detection models,this study migrated GF-5cloud detection situation to Landsat-8 situation based on the band transform.Then the cloud detection model for GF-5 is fitted based on random forest algorithm and cloud detection sample database.The results demonstrate that after the scene migration of cloud detection,the overall accuracy and classification efficiency of GF-5 cloud detection have been improved to a certain extent.The proposed approach not only effectively improves the accuracy and efficiency of the cloud detection,but also enhance the portability of the models,so as to expedite the global processing of medium-resolution satellite data.