| In recent years,with the successful deployment and operation of resource series and high-resolution series of earth observation satellites,China has realized an earth observation network with high spatial resolution,high spectral resolution,and high temporal resolution.However,nearly 60% earth surface is covered by clouds,most remote sensing imageries would inevitably be contaminated by clouds.Cloud coverage degrades the quality of satellite imagery,thus affecting imagery post-processing.Hence,it is important to quickly and accurately detect cloudage information to assess the quality of remote sensing imagery.In this thesis,we take domestic ZY-3 optical satellite remote sensing image data as the research object,explore the research on remote sensing image cloud detection from supervised and unsupervised learning,and realize the rapid and accurate cloud detection of remote sensing image.The cloud detection technology is successfully applied to the “Natural Resources Satellite Remote Sensing Cloud Service Platform” of the Ministry of Natural Resources.In general,the main contributions of this thesis can be summarized as follows:1.This thesis proposed a remote sensing image cloud detection method based on multi-scale feature extraction and object boundary refinement methods.Since the ground-objects of remote sensing images are highly complex and traditional feature extraction algorithms are difficult to extract discriminative features,this thesis proposed a feature pyramid module to extract multiscale and global contextual information for category recognition of image region under the supervised learning framework.Then,this thesis introduced a boundary refinement module to capture sharp and detailed object boundaries.Most importantly,the designed network has an encoder–decoder framework,which exploits features at multilevel layers to predict more accurate cloud detection results.Experiments show that the proposed method achieves promising cloud detection performance on ZY-3 satellite image dataset and obtains 85.61% MIo U.2.This thesis proposed a remote sensing image cloud detection method based on adaptive feature fusion and high-level semantic information remediation.Since it is difficult to identify clouds on cloud-snow coexistence images,this thesis proposed an adaptive feature fusion module to optimize the feature fusion process at different abstraction levels to improve the efficiency of feature fusion at different levels and scales under the supervised learning framework,thus improving the cloud detection accuracy.In addition,this thesis introduced a series of high-level semantic information guidance flows to remedy the semantic dilution problem and make feature maps at each level aware of the locations of the cloud objects to further improve the accuracy of cloud detection.Experiments show that the proposed method achieves excellent cloud detection results on ZY-3 satellite images with cloud-snow coexistence and obtains 85.88%MIo U.3.This thesis proposed an unsupervised domain adaptation(UDA)remote sensing image cloud detection method based on grouped features alignment and entropy minimization domain adaptation methods.Since performing visual appearance distribution alignment based on style transfer approaches cannot really reduce the distribution differences between cross-satellite datasets,this thesis proposed UDA based on featurelevel and output-level domain adaptation for cloud detection.Specifically,this thesis proposed a grouped feature alignment and an entropy minimization domain adaptation methods to reduce the domain distribution gaps between the source and target datasets,thus effectively transferring the models trained on the source labeled satellite remote sensing datasets to the target unlabeled ones.Experiments show that the proposed method achieves better cloud detection results against existing state-of-the-art UDA approaches on ZY-3 satellite images and obtains 82.73% MIo U on “GF-1→ZY-3” domain adaptation task.4.This thesis proposed an unsupervised domain adaptation remote sensing image cloud detection method based on a fine-grained feature alignment strategy.Since the traditional domain adaptation methods are difficult to effectively align feature distributions between different domains,this thesis proposed a fine-grained feature alignment method.Specifically,this thesis propose an attention-guided class-relevant feature selection mechanism to obtain local class-relevant features in the source and target domains.Then this thesis introduced the grouped feature alignment domain adaptation method to effectively reduce the local class-relevant features domain distribution gap between the source and target domains,thus improving the generalization of source domain dataset trained model on target dataset.Experimental results on ZY-3 satellite images demonstrate the effectiveness of the proposed method against existing state-ofthe-art UDA approaches.The proposed method obtains 83.09% MIo U on “GF-1→ZY-3” domain adaptation task. |
