Research On Recognition Technology Of Typical Ground-Based Cloud

Clouds play an important role in atmospheric radiative transfer and climate change. The rapid development of the whole sky imagers improves the spatial and temporal resolutions of the cloud. However, it is still a difficult task to achieve real-time cloud identification quantitatively and automatically so far. The study of the identification for ground-based cloud type is very important to the automatic observation of ground nephoscope.In this paper, the typical cloud images data are reorganized according to the visible nephograms which are obtained from digital color camera and network. Cloud formation process, the geemetry and texture features are researched based on these typical cloud images. The principle of cloud classification for ground nephoscope is researched and two kinds of measurement standards for cloud classification are presented. In the first measurement standard, clouds are classified to four basic cloud types, i.e., cirrus, undulatus, stratiform and cumuliform, according to the basic shape of the cloud. In the second measurement standard, clouds are divided to low-frequency cloud, medium-frequency cloud and high-frequency cloud, according to the cloud texture frequency. Cloud recognition algorithm and cloud classification standards are related. The first measurement standard can be selected if the cloud texture information, such as the cloud shape, color, cloud seam size features are used to the cloud image recognition. The second measurement standard can be chosen if the cloud texture frequency information of multi-scale analysis are used to the cloud image recognition.In order to reduce cloud image degradation caused by natural environment factors, such as light, rain, fog and dust and so on, the cloud image should be preprocessed before recognition to extract the useful information of the ground-based cloud as much as possible. A cloud de-illumination algorithm is researched based on Bidimensional Empirical Mode Decomposition (BEMD) combined with Closed-Form solution to natural image matting for visible ground nephogram to the de-illumination of automatic cloud type classification. This method shows pretty well de-illumination performance without losing its value information and with clear differences from background compared with algorithm based on L1 of total variation mode. Based on the natural matting of perceptual color space algorithm is proposed to overcome the problem between visible ground-based nephogram of clouds and sky background segmentation. The experimental results show that the algorithm can separate the sky and cloud effectively. A visible ground nephogram image enhancement algorithm based on adaptive fractional differential is established to extract the natural texture of visible ground-based cloud image. The results show that this method can extract image texture and edge details and simplify the process of differential order selection.The methods of ground-base cloud feature extraction are studied to effectively and steadily extract cloud image texture characteristics from the preprocessed cloud image. Based on the structure characteristics of the cloud, the gray level co-occurrence matrix algorithm can effectively extract the four texture feature of the cloud, i.e., second moments, contrast, correlation and entropy, and the BP neural network classifier can preliminarily identify cirrus, undulatus, stratiform and cumuliform. According to the significant differences between the four categories of cloud, a new BEMD and Tamura texture analysis algorithm is proposed to increase the identification of types of cloud. Texture images are decomposed into several IMFs by BEMD. Radial basis function polynomial interpolation is applied to construct the envelope. Then the number of zero-crossing, means and standard deviation in each IMFs are selected as the eigenvector for training processes. And then Tamura texture feature analysis is used to extract of the eigenvalue of orientation. Database of clouds eigenvector-sample is established by synthesizing the two normalized eigenvector. Then the images are categorized by comparing with the eigenvector of sample database calculated by the average sample method, K nearest neighbor method and support vector machine (SVM) classification algorithm. The experimental results show the recognition rate of the cumulus, altocumulus, stratocumulus, cumulonimbus, cirrostratus and cirrocumulus is obviously enhanced, especially for that of the altocumulus translucidus, cumulus humilis and cumulonimbus calvus. According to the cloud texture frequency, a novel algorithm is proposed to extract visible nephogram feature from the Hilbert spectrum of cloud image using BEMD. Cloud images are first decomposed into several Intrinsic Mode Functions (IMFs) of textural features by BEMD. The IMFs are converted from two-dimensional data to one-dimensional format, and then the Hilbert-Huang transform is performed to obtain the Hilbert spectrum and Hilbert marginal spectrum. It is shown that the Hilbert spectrum and Hilbert marginal spectrum of different types of cloud textural images can be divided into three different frequency bands by using the method of k neighbor classifier. A Contourlet and the power spectrum analysis algorithm is also proposed to capture the edges of the cloud contour, texture and geometric structure more effectively. This algorithm is proved the flexibly of multi-scale decomposition, the more direction information of the cloud image texture. The coefficient matrix from Contourlet transform of ground nephogram is calculated. The energy, mean and variance characteristics calculated from coefficient matrix are composed of the feature information. The frequency information is obtained by the power spectrum analysis, and then SVM classification of ground nephogram is used. The results indicate that contourlet transform combined with the power spectrum method is more suitable for altocumulus and stratus classification for their texture frequency difference, and the more cloud type can be effectively identified and further subdivided.After extracting ground nephogram texture features, the multiple classification algorithms, such as the average sample method, nearest neighbor, BP neural network, SVM, are applied on cloud images identification. The sample recognition rate is calculated to test the effectiveness of the cloud image recognition algorithm. To simulate the artificial thinking of weather observer realistically, the combined classifiers with dynamic selection interest operator for different clouds are researched. The combined classifier can further increase the identification of cloud types and improve the identification accuracy compared with that of a single classifier.