Research On Multiple Kernel Learning-based Cloud Detection And Analysis Technique

Cloud detection is an important task for satellite-based remote sensing images,especially helps assessing the impact of global environmental and climate changes.However, due to the impact of cloud top height and cloud thickness, the within-classvariance of cloud spectra is huge, and the traditional methods are often unable to getgood results. This thesis starts from the remote sensing images of cloud, and mainlystudies the small sample size, within-class variance, cloud features and cloud topheight and thickness in applications of multiple kernel learning based clouddetection to obtain a whole space detection for the purpose of promoting thedevelopment of kernel methods in cloud detection area.This thesis is based on multiple kernel learning theories, and its applications incloud detection and cloud top height and thickness estimations which include thefollowing three aspects in detail.First of all, we introduce some wavelengths which are sensitive to differenttypes of clouds, and based on that, we extract five groups of features。 And we findout that cloud pixel spectra are highly related to the height and thickness of cloud,so they have a huge within-class variance. And we form the cloud detection problemas an unmixing task because of mass mixed pixels brought by low spatial resolution.Then, traditional cloud detection methods are introduced include Cloud Maskand linear discriminant analysis. After that, this thesis briefly introduces the theoryof support vector machine and its development. And knowing that learning withmultiple kernels can provide better information mining and class description ability,we demonstrate it applicability on the problem of cloud detection. And this thesisprovides a multiple kernel learning based cloud detection, gives definition ofmultiple kernel Hilbert space, and in that space, the class boundary are used fordescribe cloud and background pixels. According to the ideal of unmixing, thedistances between cloud pixels and class boundaries of cloud and background inmultiple kernel Hilbert space are used for cloud fraction estimation.Finally, cloud top height and cloud thickness in multiple kernel Hilbert spaceare studied, and multiple kernel learning based cloud top height and cloud thicknessestimation methods are provided. And real remote sensing images are applied inexperiments to compare the results of CO2slicing method, support vector regressionmethods, and single kernel method with same form.