Research On Cloud Detection Algorithm Based On Deep Learning

The rapid progress of remote sensing technology has made the country safer,the society more civilized and the life more convenient;from military field,land planning to urban traffic and weather disaster warning,remote sensing technology can be said to have penetrated into every aspect,human society has been highly civilized,and remote sensing technology has played an important role.However,there are still some problems to be solved in the use of remote sensing data,and clouds are one of the most important effects,which put blinders on human satellites and make remote sensing images not to show the value of the whole picture.According to the measurement,more than half of the earth’s surface is covered by clouds all year round,and these clouds come in all shapes and sizes,but they all block the effective feature information,and the bandwidth that can be used for data transmission between satellites and the ground is extremely limited.If cloud rejection is performed on remote sensing images before satellite data transmission,this will make air-ground data transmission much more efficient.Therefore,in this paper,we will study remote sensing images and design an effective cloud detection algorithm.There are various forms of clouds in remote sensing images,such as thick and dense clouds,fish scale clouds,fine broken clouds,etc.Meanwhile,some high-bright features,such as white building clusters,highly reflective water surface,snow and ice in high mountains,are extremely similar to clouds,which make the task of cloud detection in remote sensing images complicated.To achieve effective separation of clouds and features,traditional methods need to obtain thresholds through empirical or complex data analysis,or select suitable ones among a large number of features for segmentation,which are inefficient and have low detection accuracy.This paper selects deep learning algorithm and relies on its powerful characterization ability to design a suitable and efficient cloud detection algorithm.The main contents and innovations of this paper are as follows:The proposed convolutional neural network for cloud detection based on codec structure merges the features of each layer at the coding end to the decoding end using jumping connection lines;in order to enhance the information of the cloud region of interest while suppressing the useless information of the features,the spatial attention module and the channel attention module are designed,in which the spatial attention module is based on the non-local idea to enhance the features of the cloud region of interest and is placed at the last The channel attention adopts the feature of multi-scale information fusion and fuses the information at the coding end and puts it on each layer of features at the decoding end;a probabilistic upsampling module based on a priori information is proposed to improve the problem that the detection of edge regions of clouds is not continuous enough.In order to make full use of the rich spectral and spatial information,a dense connected strategy based spectral-spatial feature extraction module is designed,which effectively realizes the independent extraction of spectral and spatial information,at the same time,in order to enhance the effective information and suppress the useless information,a self-attentive-based spatial and channel In order to enhance the effective information and suppress the useless information,the spatial and channel attentions based on self-attentiveness are designed and added after the spectral information extraction module and the spatial information extraction module respectively,finally,the contextual dynamic convolution module is designed to make the network adjust the convolution kernel parameters adaptively and enhance the characterization ability of the network.Finally,the performance of the cloud detection algorithm for remote sensing images proposed in this paper is proved to be better than other algorithms through experiments.