Research On Remote Sensing Image Cloud Detection Algorithm Based On Deep Learning

Cloud cover plays an important role in met eorological elements,and meteorologists can analyze and predict climate further by observing changes in cloud cover.Additionally,more than 50% of the Earth’s surface is covered by clouds,and cloud cover affects many applications based on remote sensing images.Consequently,cloud detection is crucial in remote sensing image analysis and pre-processing.With the rapid development of deep learning technology,deep learning cloud detection algorithms have significantly outperformed traditional algorithms.However,due to the diversity and complexity of remote sensing images,existing cloud detection algorithms suffer from problems such as missed detection and false detection,which influence the accuracy of downstream tasks such as remote sensing image clas sification and segmentation.Therefore,studying cloud detection algorithms based on deep learning is of great theoretical significance and practical value.The research work presented in this thesis is as follows:(1)A feature-aware aggregation network(FAANet)for remote sensing image cloud detection is proposed in this thesis,specifically targeting the convolutional neural networks(CNNs)feature fusion process.Unlike previous approaches that overlook the contribution and differences of features,FAA Net leverages feature enhancement and feature aggregation to maximize the valuable features for cloud detection.The feature enhancement module(FEM)consists of two components: the high-level feature enhancement module(HFEM)enhances high-level semantic features to complement cloud features lost during the recovery of high-resolution feature maps,and the lowlevel feature enhancement module(LFEM)enhances spatially detailed features of remotely sensed images to further refine the predicted cloud images.The feature aggregation module(FAM)selectively aggregates the coding module,and decoding module,and enhances semantic features using a new aggregation mechanism to supplement features diluted by upsampling,thereby improving the integrity of predicted cloud images.The efficacy of the proposed approach is validated through extensive comparison and ablation experiments conducted on three open-source cloud detection datasets and the higher spatial resolution Landsat8 cloud detection dataset established in this thesis.Results demonstrate that the FAANet achieves competitive performance across different evaluation metrics.(2)To address the limitations of convolutional neural networks(CNNs)in reasoning about inter-feature relationships and the problem of semantic gaps during the multilevel feature fusion process,a feature interaction graph convolutional network(Cloud-Graph)algorithm for remote sensing image cloud detection is proposed in this thesis.The algorithm comprises three main components: re mote sensing image feature extraction,feature interaction graph inference,and high-resolution feature recovery.The proposed feature interaction graph reasoning(FIGR)module first fully interacts with low-level spatial detail features and high-level semantic features using an attention mechanism and then uses a residual graph convolutional network to infer the higher-order relationships between the two features.The algorithm effectively alleviates the semantic gap problem during cross-level feature fusion,while the fused features aggregate valuable detail and semantic information.Experimental evaluations on the open-source 38-Cloud and SPARCS datasets,as well as the self-built CHLandsat8 dataset in this thesis,demonstrate that the proposed algorithm better detects clouds in remote sensing images with complex cloud shapes,sizes,thicknesses,and cloud-snow coexistence.(3)To address the limitations of the convolutional neural networks(CNNs)when modeling long-range relational contexts and the bottl eneck problem of cloud detection accuracy,a parallel branch network fused Transformers and CNNs(Trans CNet)for remote sensing image cloud detection is proposed in this thesis.The algorithm combines the advantages of Transformer and CNN,including a Tran sformer branch that learns global contextual features and a CNN branch that learns local spatial details,and the elements of the two branches are fused complementarily at each encoder stage.To effectively fuse the different levels of features from the tw o branches,a novel feature aggregation module continues to be designed.Finally,to obtain a clearer cloud layer,elements from the feature aggregation module are used to complement the features sampled missing on the decoding block,while the fused featu res are enhanced using standard channel attention mechanisms.Extensive comparison experiments and ablation experiments are conducted on both publicly available datasets and the self-built cloud detection dataset in this thesis,and the designed method ach ieves competitive performance under different evaluation metrics.