Research On Cloud Detection Algorithm For Sensed Satellite Imagery Based On Deep Learning

With the rapid development of satellite remote sensing technology in China,a large number of high-resolution remote sensing images have been applied to many fields,including agricultural monitoring,environmental protection,and urban planning.However,according to monitoring data from the International Satellite Cloud Climatology Project,more than65% of the world’s area is covered by clouds.The images taken by remote sensing satellites are inevitably affected by clouds.And due to the reflection of clouds on different spectral bands,most remote sensing satellite cameras do not have spectral bands that can penetrate the clouds,so the features in the satellite images taken will be blocked by the clouds.The problem of occlusion formed by the cloud layer in the image will have a greater impact on the research of target recognition,image fusion,and feature classification using remote sensing images.Therefore,cloud detection is a key preprocessing step in the field of remote sensing images,and it is also a hot issue in remote sensing images.Remote sensing image cloud detection can be regarded as an image segmentation problem.In the field of image segmentation,deep learning algorithms have far outperformed traditional segmentation algorithms.In order to achieve better results,traditional methods for cloud detection of remote sensing images usually use a variety of complex image color transformation methods and texture feature extraction methods to optimize the algorithm based on a large number of prior conditions.Sometimes it is even necessary to manually set parameters such as relevant thresholds according to the different types of underlying surface.The convolutional neural network model in deep learning algorithms can autonomously mine a large number of shallow and deep abstract features from the image,and expand and transform the data to obtain better algorithm performance.Therefore,this paper is based on deep learning algorithms to study the cloud detection in high-resolution remote sensing images.The main content and innovations of this paper are as follows:A cloud detection network based on dense connections and encoder-decoder structure is proposed.Cloud detection is end-to-end image segmentation,that is,the width and height of the output image of the network are the same as the input,and the output is a binary map containing the pixel classification results.Therefore,an end-to-end encoder-decoder network structure framework is adopted.At the same time,in order to fully extract and transform the features in high-resolution images,this paper adds the dense connection module designed in this paper to the encoding and decoding ends,and adds a jump connection between the encoding and decoding ends.These can enhance the network’s reuse of shallow features and the transformation of deep abstract features,and improve the information interpretation capability of the decoder.On the basis of the codec network framework with skip connections in the first part,a new network called based on Gabor feature extraction module and channel attention module(NGCA)is proposed from the perspective of accurate information enhancement.Because the texture details in high-resolution remote sensing images are rich,and texture features are important information in cloud detection,combining the advantages of Gabor transform to extract texture features in multiple scales and directions,the Gabor feature extraction module is designed.This module is added to the encoder,thereby guiding the network to learn the texture structure of the image,and strengthening the network’s attention to the texture details.Because the feature maps generated by the decoder contain different types of information,some feature maps contain more cloudy area information and some contain more non-cloud area feature information.Therefore,by adding the channel attention module,the information of the encoder is introduced at the decoder to enhance the important information in the deep abstract features,and the useless information is weakened,further improving the network performance through quasi-information enhancement mechanisms.The experiments show that the cloud detection network proposed in this paper has improved the accuracy rate by 1.04 % and 0.51 % and reduced the false detection rate by 4.51 %and 0.72 %,respectively,compared to Deeplabv3+ and RS-Net,a cloud detection network based on encoder-decoder architecture.