High-dimensional LBP Feature Selection Method For Meteorological Cloud Image Classification

Cloud shape(cloud type)is a very important parameter for ground-based meteorological cloud image observation.It refers to the type and shape of clouds.The correct recognition of cloud type is very important for the understanding and analysis of climatic conditions,numerical weather prediction and atmospheric circulation patterns.However,the structure,contour,texture and color of the cloud in the image are uncertain,and they are all easily affected by the light,shooting environment,and shooting location.This makes the study of cloud type recognition for ground-based meteorological cloud image face greater challenges.Specifically,in the cloud type recognition tasks of ground-based meteorological cloud images,useful features are usually extracted from the cloud images first,and then cloud type recognition tasks are performed based on these features with the help of classifiers.Feature extraction and selection is the most important step,because the quality of feature selection directly affects the classification results of ground-based meteorological cloud images.What's more,the feature set extracted from ground-based meteorological cloud images tends to have high dimensionality,that is,the dimensionality of the feature set is much higher than the number of samples.For example,the dimension of the feature set extracted by the Local Binary Pattern(LBP),which is widely used in ground-based meteorological cloud image analysis,is at least 1-2 orders of magnitude higher than the number of cloud images.Therefore,it is very important to study the selection of high-dimensional LBP feature in the analysis of ground-based meteorological cloud image.The current methods of the selection of high-dimensional LBP feature set mainly include LTP,CLBP,DLBP,Sa LBP,SLBP etc.However,these methods only use the information of the feature itself rather than the category information of the cloud in the feature selection process.Therefore,the information between LBP feature and cloud categorie is fully considered in this paper,and a new F-like statistic is constructed,which uses the idea of forward search to conduct the selection of high-dimensional feature.The main contributions of this paper include the following points:(1)A new method of the selection of high-dimensional LBP feature based on mutual information F-like statistics is proposed.With the help of the mutual information between feature and categorie variable,the useful information of ground-based meteorological cloud image recognition is fully extracted,which makes up for the defect of other methods of LBP that only use the information of the cloud image itself but not the category information of the cloud image.At the same time,the redundant information between high-dimensional features is considered,so that better features can be selected.(2)It is theoretically proved that the proposed method of the feature selection of F-like statistics based on mutual information has the consistency of selection,that is,the optimal feature set selected based on the proposed method under finite samples will also be optimal when the sample tends to infinity,which provides a theoretical guarantee for the practical application of the proposed method.(3)Experiments on three real ground-based meteorological cloud image databases show that the performance of the method proposed in this paper is superior to the performance of rotating LBP,consistent rotating LBP,CLBP,DLBP,and maximum correlation based on mutual information.Experiments are based on five classifiers of Support Vector Machine,Naive Bayes,k-Nearest Neighbor,Neural Network,and Multiclass Mogistic Regression,and two-fold,five-fold,and ten-fold cross-validation.