Remote Sensing Image Cloud Detection Algorithm Based On Deep Adversarial Network

With the vigorous development of remote sensing technology,remote sensing image sensors are more sensitive,it is easier to obtain remote sensing images,and the types of remote sensing images are more abundant.At the same time,multispectral and hyperspectral remote sensing images are becoming more and more mainstream.Scholars can use multispectral and hyperspectral remote sensing images with rich spectral information for research in various fields.Because a large amount of cloud cover in the atmosphere will affect the subsequent image processing,the cloud detection of remote sensing images has also received extensive attention from scholars.However,the complexity of multispectral and hyperspectral remote sensing images poses severe challenges to current cloud detection.The main problems are as follows:(1)High-dimensional remote sensing images face the issues of large data volume and insufficient training samples,which affect the accuracy of cloud detection.(2)There is redundancy and noise information in the spectral information of multispectral and hyper-spectral images,which leads to unsatisfactory effect of high-dimensional image cloud detection.(3)Existing cloud detection algorithms are highly complex and do not meet the real-time processing requirements on remote sensing satellite platforms.Therefore,it is of great significance and application value to study how to improve the accuracy of cloud target detection in multi-dimensional remote sensing images and solve the problem of real-time processing of cloud detection on the satellite.The main research contents and innovations of this paper are as follows:Based on cloud targets can be modeled as low-rank sparse values,a confrontational autoencoder network model is designed for multispectral and hyperspectral images to extract cloud target features,and the network is forced to perform distinguished feature learning to obtain the original input and the residual between two backgrounds in the deep feature space.A compact representation of the original image is obtained by the potentially adversarial learning constraint encoder.Since the background pixels are the majority of training samples,the decoder reconstructs the background pixels accurately,and the image discriminator is used to prevent the generalization of over-class features due to potential adversarial learning.In order to further highlight the background information in the deep feature space,a variational autoencoder is introduced to extract background features.Finally,the residuals are used in the deep feature space to highlight cloud features and suppress background features.For the extracted cloud feature maps,a cloud target detector based on structure tensor adaptive fusion,attribute filtering and guided filter combination is designed to highlight the highlight features of cloud targets while maintaining the edge of cloud targets and improving the detection rate of thin clouds.This paper proposed an iterative detection algorithm,according to the detection results,which know the original features and refuse to achieve the purpose of suppressing the background.Meanwhile,the algorithm use the calculation of the similarity of two adjacent detection results to make the algorithm stop iterating intelligently.In response to the needs of remote sensing satellites for real-time processing,the above algorithms are optimized,and then the optimization plan for on-board real-time processing is proposed.After a deep understanding of the network principle,a solution is designed for ground training and extracting in real time on the satellite.The solution uses low-complexity morphological filtering instead of attribute filtering,and optimizes the corrosion and expansion operations of morphological filtering,and optimizes the mean filtering module of the guided filter.Through the above solution,the processing speed of the algorithm is greatly improved and resource consumption is greatly reduced.The experimental results show that the cloud detection OA value of the algorithm proposed in this paper is above 0.95 on the Landsat8 dataset and above 0.95 on the GF1-WFV dataset.The processing speed of this algorithm is the fastest among all the comparison algorithms.In summary,the algorithms and schemes proposed in this paper can be used as the theoretical and application basis for remote sensing satellite cloud target detection.