The Domain Adaptation Research For Remote Sensing Imagery Cloud Detection Based On Deep Learning

Clouds significantly reduce data availability while increasing labor,transmission,storage,and computing costs.Therefore,an effective cloud detection is a crucial step in the remote sensing image preprocess in order to indicate the quality of the available images and minimize resource waste.In the last ten years,deep learning has advanced quickly in the research and application of remote sensing.This technique has been used by numerous research teams to perform cloud detection for remote sensing images.However,due to the limitation of feature extraction mode and data domain offset the current methods still have the following drawbacks:(1)The previous cloud detection networks were difficult to capture the effective global features and prone to loss significant fine grain information in very-high resolution image patches of snow areas.It results thin cloud missing detection and low-level error detection.(2)The generalization ability of the cloud detection model will have a significant impact if the distribution of test data differs significantly from the training data.In this study,deep learning technology serves as the foundation,cloud detection neural networks and domain adaptive technology serve as the major research paths.The following are the primary research contents and accomplishments:(1)Study on multi-task strategy and reconfigurable mechanism for cloud detection neural networkThis paper proposes a multi-task driven and reconfigurable network(MTDR-Net)and creates a very-high resolution snow region cloud detection dataset(Cloud S26)to better perform cloud detection in high-resolution imagery.In MTDR-Net,the highresolution backbone with multiple projection heads module facilitates global information interaction,the multi-level granularity feature representation,and feature sharing between pixel-level and superpixel-level segmentation tasks.The reparametrizable multi-scale feature fusion module is used to capture the local multi-scale cloud features in the training stage and simplify losslessly the structure in the testing stage.The lightweight and adaptive feature fusion module is employed to reconstruct meaningful cloud features.The multi-task gradient flow guidance module can provide unbiased guidance on dividing gradient flow between the multi-tasks.The experimental results show that the MTDR-Net obtains the greatest accuracy in the comparison methods,which due to its anti-interference capacity for confusing ground objects and its excellent thin cloud boundary extraction ability.In addition,MTDR-Net has the fewer parameters,forward-reasoning floating-point operations,memory footprint,and memory access volume.With the coarse-to-fine strategy,MTDR-Net can be further sped up by avoiding intensive inference resource consumption in cloud-free/all-cloud scenes.(2)Research on cloud detection domain adaptation based on image-level simulation and feature-level alignment optimizationTo lessen the accuracy loss of cloud detection neural networks on various data distribution domains,a two-stage domain adaption based on image and feature levels(TDAIF)cloud detection framework is constructed in this study.At the image level,TDAIF successfully fuses source domain foreground information and target domain background information by pseudo target domain data generator.It is utilized to the model explore the invariance semantic knowledge of the target domain.At the feature level,utilizing adversarial domain discrimination and self-ensembling consistency strategy,TDAIF implicitly handles global feature alignment and decision boundarylocal feature optimization.Finally,through combining the image and feature level processing,the influence of image radiometric diversity and scale divergence is weakened,and the adaptive generalization capability of network for joint distribution of domain offsets is enhanced.The experimental results based on cross-space-time and cross-sensor remote sensing datasets show that the TDAIF significantly reduces the cross-domain accuracy loss of cloud detection neural network.Furthermore,TDAIF obtains the highest cloud detection accuracy among the deep transfer learning methods.In addition,in the visual results on cross-domain remote sensing images,TDAIF shows extremely few low-level cloud misclassifications and missed detection.