Development of a support vector machine based cloud detection method for MODIS with the adjustability to various conditions

Common requirements for cloud detection methods including the adjustability with respect to incorrect results are clarified, and a method is proposed that satisfies the requirements by applying the support vector machine (SVM). Because the conditions of clouds and Earth's surfaces vary widely, incorrect results in actual cloud detection operations are unavoidable. Cloud detection methods therefore should be adjustable to easily reduce the frequency of incorrect results under certain conditions, without causing new incorrect results under other conditions. Cloud detection methods are also required to resolve a characteristic issue: the boundary between clear-sky and cloudy-sky areas in nature is vague, because the density of the cloud particles continuously varies. This vagueness makes the cloud definition subjective. Furthermore, the training dataset preparation for machine learning should avoid circular arguments. The SVM learning is generally less likely to result in overfitting: this study suggests that only typical data are sufficient for the SVM training dataset. By incorporating the discriminant analysis (DA), it is possible to subjectively determine the definition of typical cloudy and clear sky and to obtain typical cloud data without direct cloud detection. In an approach to adjust the classifier, data typical of certain conditions that lead to incorrect results are added to the training dataset. In this study, an adjustment procedure is proposed, which quantitatively judges, whether an addition is actually effective for reduction of the frequency of incorrect results. Another approach for the adjustment is improving feature space used for cloud detection. Indices as quantitative guidance to estimate whether an addition or elimination of a feature actually reduces the frequency of incorrect results can be obtained from the analysis of the support vectors. The cloud detection method incorporating the SVM is therefore able to integrate practical adjustment procedures. Applications of this method to Moderate Resolution Imaging Spectroradiometer (MODIS) data demonstrate that the concept of the method satisfies the requirements and the adjustability to various conditions can be realized.