The Research Of Cloud Detection And Cloud Removal For Remote Sensing Images Based On Reference Image Information Fusion

Remote sensing images are the only data resource that can observe the earth in a large range.They are widely used in agriculture,military,urban planning and other fields.However,the remote sensing image is inevitably polluted by the cloud,haze and other atmospheric environment,resulting in the loss of image information.These images can not meet subsequent practical applications,which resulting in a huge waste of data resources.Therefore,cloud detection and cloud or haze removal for remote sensing images are effective to improve the utilization rate.In this paper,cloud detection algorithms and cloud or haze removal algorithms are studied.In terms of cloud detection,because different remote sensing satellite sensors have different spectral distribution,the cloud detection algorithms based on multi-band are for some specific remote sensing sensors.The detection algorithms based on machine learning and deep learning have good applicability,but the complexity is high and the accuracy of haze detection is low.In terms of cloud or haze removal,it is a challenge to remove haze or cloud while keeping the information of clean areas unchanged.Since the remote sensing images have the characteristics of multi-temporal,the information of these multi-temporal images is complementary,which can be applied for cloud free image reconstruction.Therefore,this paper mainly studies the cloud detection and cloud removal methods based on multi-temporal remote sensing image information fusion.The main work and innovation of this paper are as follows:(1)A cloud detection algorithm based on reference image fusion is proposed.Firstly,the dark channel images of the target image and the reference image are obtained,then the initial thick cloud detection results are obtained by the threshold method,and are optimized by the guided filter.After the detection of thick cloud,the pixels in other regions are extracted by kernel principle component analysis(KPCA),and finally the thin cloud is detected by threshold method.The evaluation results ofthe correct detection accuracy,the missed detection accuracy and the overall detection accuracy show that the algorithm can effectively detect the thick cloud and the thin cloud,and reduce the false detection rate caused by the interference of the highlighted objects.(2)A thin cloud removal algorithm based on reference image information fusion is proposed.Firstly,the difference between the thin cloud image and the reference image is used to estimate the atmospheric transmittance,and the initial cloud removal image is obtained by the thin cloud removal model.Then,the initial cloud removal image and the dark channel image of the original image are used as input to obtain the thin cloud distribution mask.Finally,the wavelet image fusion algorithm is used to optimize the cloud removal results and get the final cloud removal results.The haze effect index and structure similarity are used to analyze the removal results of thin cloud.It shows that the method can get a good cloud removal effect and ensure the fidelity of the cloudless areas,even if there is obvious difference between the reference image and the thin cloud image.(3)A cloud removal algorithm based on selective multi-source total variation model is proposed.In order to overcome the difficulty of obtaining a completely cloudless reference image,our method select one of the multi-temporal remote sensing images as target image for reconstruction.Firstly,the brightness of multi temporal remote sensing image is corrected to reduce the brightness difference between images.Then the image reconstruction is completed based on the selection of multi-source total variation model.Finally,Poisson image optimization is carried out to solve the problem of luminance information.The evaluation of the average gradient and structure similarity of the cloud removal results shows that the algorithm effectively solves the problem that the selected target image cannot be reconstructed by selecting multi-source total variation model,and ensures the image quality of the thick cloud removal results.Through the experiments of remote sensing images with different terrain types,and compared with the results of other algorithms,the cloud detection algorithm and cloud removal algorithm proposed in this paper can yield better results.It verifies the effectiveness and feasibility of the algorithms in this paper.