Classification of Whole Sky Infrared Cloud Image Using Compressive Sensing

Cloud type, as an important macroeconomic parameter in cloud detection, plays a mean role in weather forecasting, field meteorological service, aerospace and climate researches. Automatic identification of cloud types is not efficiently resolved. Cloud shapes, texture, color, contour, range, process of change and some other features are used for manual cloud classification, but it is hard to find a nice way to extract effective features for automatic identification. Particularly, infrared images provide less resolution and less color information.A new method is proposed to classify cloud images obtained from the whole sky infrared cloud measuring system (WSIRCMS) from compressive sensing (CS). Firstly, a redundant dictionary is constructed with typical cloud samples. In order to reduce the computational complexity and computing time, principal component analysis (PCA) and down-sampling is applied to dimension reduction in building up redundant dictionary. It's found that classification results tend to be stable and suitable when the feature contribution rate is more than 95% in PCA or at 16-time down-sampling. Secondly, the optimal solution of paradigm is solved using gradient projection for sparse reconstruction (GPSR) and orthogonal matching pursuit (OMP) algorithms. Sparse algorithm has a certain influence on classification results. There are some negative sparse solutions in GPSR and OMP algorithms, and through the analysis, when the proportion of negative sparse solution is more than 46%, the classification of residual method is prone to error. Sparse solution may be wrong if the incoherence of different type cannot be guaranteed in establishing redundant dictionary, and the dimension reduction may especially increase the correlation. If the cloud texture, structure feature can be kept in process of dimension reduction and one-dimensional treatment and establishing redundant dictionary is complete, it probably makes better sparse solution. Finally, the residual method and sparse proportion method are used to discriminate cloud types. According to experimental results, it's found that the spare-proportion of wave cloud misclassified as cumuliform is less than cumuliform, and for the wave cloud misclassified as stratus cloud or cirrus, its spare-proportion of stratus and cirrus type is small. By combining two discriminated methods, two greatest sparse proportion types are selected and then the small residual is analyzed. Classification accuracy of wave, cumuliform, cirrus cloud is improved.Using compress sensing theory in cloud classification avoids the feature extraction process, and provides a new way for the automatic identification of infrared cloud images. With this method, the recognition rate of waveform, stratiform, cumuliform, cirrus and clear sky reaches 75%, 91%, 70%, 85% and 93%, respectively, with the average accuracy up to 82.8%.