Retrieval And Removal Of Thin Clouds And Haze From Multispectral Remote Sensing Imagery

Spaceborne optical remote sensors are inevitably affected by the clouds in the atmosphere when acquiring the information of ground objects.The existence of cloud or haze has greatly affected optical remote sensing data,which can not meet the requirements of data quality and quantity in remote sensing research and application,especially in quantitative remote sensing and time series analysis.Studying cloud detection and cloud removal algorithms that are fast and efficient can significantly improve the utilization rate of existing optical data and make up for the shortage of data in remote sensing scientific research and application.Cloud detection and removal have been studied for decades.However,the cloud detection and removal in multispectral remote sensing data,especially for thin clouds,is still a difficult problem and challenge faced by remote sensing,geography,and other disciplines due to the uncertainty of cloud in space and time.For cloud detection,the level of accuracy for thick clouds detection is high,while the detection of thin clouds is quit difficult,and often depends on the specially designed cirrus band.Cloud removal is usually limited by the requirements of the data,the accuracy of the result,the applicability of spectrum and other factors.In addition,cloud detection is usually treated as a binary classification problem,while cloud removal is a ill-conditioned equation solving problem.The relationship between cloud detection and cloud removal is overlooked in traditional researches.Therefore,this dissertation mainly focus on thin clouds.Based on the physical characteristics,the electromagnetic radiation mechanism of thin clouds has been studied and the remote sensing imaging model under the condition of thin cloud has been discussed.Then,the theoretical framework of cloud parameters retrival and thin cloud removal has been constructed,which can reduce the demand for data and improve the accuracy and applicability of the algorithms,and it is suitable for multi-source remote sensing data.Finally,the scientific problem that optical data are easily affected by atmospheric clouds has been sovled,and the requirements of data consistency in subsequent researches can be meet.The main works of this dissertation are shown as follows:(1)A thin cloud removal algorithm for 400-760 nm spectrum has been proposed to solve the problem that thin cloud removal algorithms with high-precision often rely on specially designed cirrus bands.Based on the physical characteristics of thin clouds and the traditional atmospheric radiation transfer model,the remote sensing imaging model under the condition of thin clouds has been established and the unified theoretical framework of thin cloud parameter retrival and removal has been constructed.Three assumptions have been made and verified in the visible and near infrared bands,which are the most frequently carried by optical satellites.The mathematical expression for removing thin clouds in any visible band has been derived and six unknown parameters have been solved.Compaired with state-of-the-art approaches,the algorithm which has wide applicability and high robustness is independent with auxiliary data.Thus,the algorithm is suitable for most current optical remote sensing satellites data which have visible and near infrared bands.(2)Considering that the thin cloud detection is difficult and the studies are mainly focus on cirrus detection and binary masks,an algorithm of retrieving the TOA reflectance of clouds has been proposed for detecting various types of clouds.Under the unified theoretical framework of thin cloud parameters retrival and removal developed above,the expression of TOA reflectance retrieval of thin clouds has been derived by using empirical relationships and multi-channel radiative transfer equations.By using the adjacent similarity and the different spectral responses of thin clouds in different short-wave bands,the method of parameters solution has been simplified,and requirements of the data have been reduced.The algorithm does not depend on auxiliary data and has been verified in multispectral data(e.g.,Landsat-7,Landsat-8)and hyperspectral data(e.g.,AVIRIS).Quantitative results show that cloud masks generated by the developed algorithm is more sensitive to thin clouds.In addition,the algorithm provides another way to obtain the reflectance of thin clouds and haze instead of the specially designed sensors.The algorithm has wider applicability since it is suitable for various types of thin clouds and haze and it does not depend on the conditions of the atmosphere.The developed algorithm provides a new approach for thin clouds detection and a new way for removal.(3)On the basis of the above researches,a thin cloud removal algorithm with highprecision for 400-2300 nm spectrum has been proposed,which can deal with all kinds of thin clouds and has wider applicability.Instead of cirrus detecton band,the TOA reflectance of thin clouds has been retrieved by the linear transformation,which can overcome the dependence on auxiliary data,and avoid the missing detection of other types of thin clouds and noises by cirrus detection band.Then,each band has been combined with the retrieved thin cloud band into pairs,and the cloud component and noncloud component have been extracted from each pair of data by independent component analysis(ICA),which can ensure the uniqueness of cloud component and non-cloud component and improve the separation accuracy of mixed signals.Finally,based on the prior knowledge that the retrieved thin cloud image should be zero after cloud removal,a new cloud component after cloud removal has been generated,and the cloud influence in each band has been removed by inverse ICA transformation.The results show that the algorithm can not only reduce the requirements for data,but also can deal with the influence of various types of thin clouds in diffierent bands,which is more practical.