Research On Ground-based Visible Cloud Image Classification Method Based On KNN Algorithm

Cloud plays an important role in the meteorological research, and it is one of themost important factors of Earth’s energy balance and hydrological cycle. In order toactualize the automatic ground-based observation of clouds, automatic classification ofcloud image is a difficult problem needed to be solved. The observation of cloud isdepended visual, and the visual observation is based on the cloud classification schemeof “china cloud atlas”. When the cloud is classified by the machinery, the erroneousclassification is serious by the standard of visual observation.There is not an object standard in cloud classification, and the existing cloudclassification is preliminary discussion. This paper presents two broad classificationschemes of cloud by the cloud of physical mechanism and shape, one contains cumulus,stratus and cirrus, another contains cumulus, stratus and undulates. The clear sky as aspecial sky condition is considered as a separate category in the scheme.The first step is extracting cloud image characteristics for cloud classification. Byanalyzing the texture features, color features and shape features of four different skyconditions (cumulus, stratus, cirrus and clear sky), this paper introduced the extractionalgorithms of the three kinds of features in detail. The texture features of cloud imagewere extracted by using gray-level co-occurrence matrix (GLCM) and Tamura texture.The color features of cloud image were extracted by using color moment. The shapefeatures of cloud image were extracted by using moment invariants.After extracting the characteristics of the cloud images, the classifier is selected.This paper introduces several classifiers in cloud classification. By comparing theseclassifiers, the K-Nearest-Neighbor classification algorithm (K-Nearest-Neighbor, KNN)is selected in this paper because of its high performance in solving complex issues,simplicity of implementation and low computational complexity, and it is one of theeffective and simple algorithms.By analyzing the first cloud classification scheme (cumulus, stratus, cirrus and clear sky) and the second cloud classification scheme (cumulus, stratus, undulates andclear sky) texture features, color features and shape features of four different skyconditions, we selected the K-Nearest-Neighbor classification algorithm in the study. Inthis paper, we used the KNN classification algorithm in8different K values anddifferent features combining to recognize the four types of sky conditions. there areseven feature combination(texture features, color features and shape features alone, anytwo of them combined, these features together) The result of the two cloud classificationscheme has the following conclusions:(1) Using texture features，color features and shape features together could getbetter performance than using these features alone or any two of them combined.(2) When the parameter K is set to7and all of21characteristic parameters areapplied, the KNN classification algorithm can obtain the best performance, the accurateidentification rate of cumulus, stratus, cirrus and clear sky are91.1%、74.4%、70.0%and100.0%respectively. The average accurate rate is83.9%. And the accurateidentification rate of cumulus, stratus, cirrus and clear sky are91.1%、74.4%、70.0%and100.0%respectively. The average accurate rate is83.9%.After comparing two cloud classification results. By optimizing the first classifierscheme of feature combination (the average of83.9%),21characteristic parameters areextracted in the first cloud classification. Some characteristics are redundant, theoptimizing the feature combination is to reduce the impact of these redundant featuresfor the recognition results, the average of recognition is improved after optimizingfeature combination. When the parameter K was set to7, the accurate identification rateare90.0%、80.0%、72.2%and100.0%respectively, with an average of85.6%. Throughthe optimization of the feature combination, the recognition accuracy and speed areimproved, the superfluous feature is reduced.