Cloud Detection Based On Deep Learning In Remote Sensing Image

The development of optical remote sensing satellites has provided great convenience for human observation of the earth.The satellites have many advantages,and their images have a wide range,high resolution,and cover rich data,so they are widely used in the fields of earth observation,geological exploration,ocean monitoring,urban planning,earthquake monitoring,etc.However,the problem is that the presence of large-scale clouds in optical remote sensing images causes serious interference to the interpretation of ground objects in subsequent images.The sky is covered by clouds all the time,and according to the survey,about 66% of the global area is occupied by clouds,and the clouds will reflect most of the signals from remote sensing satellites,so it is difficult to obtain the complete information of the ground objects,resulting in the lack of image information,which is difficult to use in many cases.Therefore,in order to improve the quality of optical remote sensing satellite images,cloud detection has an important research significance.The cloud detection method based on convolutional neural network can automatically extract effective features by stacking multiple convolutional and pooling layers,and through continuous convolution and downsampling operations,the network can obtain features such as cloud contour,texture and distribution,thus obtaining higher accuracy than traditional cloud detection methods.However,the cloud detection target morphology varies greatly,thin cloud features spectral reflectance is small,easy to be confused with the ground objects and its boundary features are not obvious with less features;small cloud pixel distribution is more scattered and the number of pixels is less,all of which are easy to produce missed detection in ordinary convolutional neural network detection.High-resolution optical remote sensing images can provide many target information for cloud detection,such as spectral information,spatial distribution information,transmittance information,etc.However,because there is less research on thin cloud detection based on deep learning,the number of thin cloud samples in the training set is small,which cannot support the network for the full extraction of thin cloud features.To address the above problems,combined with the CNN-Transformer network structure,it has greater advantages over traditional CNN networks in terms of long-range dependencies and global feature capture,and can retain more local detail information.Therefore,this paper mainly adopts the CNNTransformer-based cloud detection network structure for systematic research.The main contents of this paper and the innovation points are as follows:(1)A cloud detection method based on multi-scale feature fusion and attention mechanism is proposed,which mainly focuses on the problem of cloud detection in traditional convolutional neural networks,which has fewer features and poor edge detection,and is easy to produce missed detection.The multiscale feature fusion method is designed to make up for the semantic information difference between shallow features and deep features directly.Compared with the attention based detection algorithm,this method has a 1.9% improvement in accuracy on the GF-1 public dataset.(2)A cloud detection data generation method based on scene synthesis is designed.Mainly for the current problem of thin cloud samples in thin cloud detection dataset based on deep learning,the Gabor filtering method is firstly used to extract the features related to cloud targets at each scale and direction,and then the extraction capability of thin cloud edge information is enhanced by image fusion.The multi-channel region growth algorithm is designed to grow and fuse each channel of the image after the filtering algorithm,and the segmentation results are superimposed on different scenes to expand the thin cloud samples.(3)A CNN-Transformer-based cloud detection framework is proposed.To address the problem of the perceptual field limitation of CNN networks,in order to improve CNNs that require many layers to obtain abstract global information,this paper introduces the CNN-Transformer network structure so that it can retain more local information at the edges on top of ensuring the totality of cloud detection.Meanwhile,for the characteristics of small and scattered pixels of small clouds,this paper proposes a cloud mask weighted loss function,which constructs weights according to the pixels classified as clouds,and improves the attention of the network for small pixel cloud targets by increasing the loss weights of small pixel cloud targets.It is experimentally shown that the cloud detection method based on the CNN-Transformer framework improves the detection accuracy by 1.7% compared to the algorithm only using the attention mechanism.