Classification And Segmentation Of Ground-based Cloud Images Based On Deep Learning

This paper uses the four published ground-based cloud datasets:Singapore Whole-sky Imaging Categories(SWIMCAT)dataset,SWIMSEG(Singapore Whole Sky Image Segmen-tation)dataset,BENCHMARK dataset and Cirrus Cumulus Stratus Nimbus(CCSN)dataset Deep learning-based methods for the problems of ground-based cloud classification and cloud segmentation have been studied.The main conclusions are as follows(1)For the classification of ground-based cloud images in the meteorological field,a new convolutional neural network model was proposed,called CloudNet,for ground-based cloud classification.A ground-based cloud dataset was established,called CCSN,which is a dataset including 10 cloud types under meteorological standards.The number of cloud images in the CCSN dataset is three times that of the previous dataset.For the first time,the contrails gener-ated by the aircraft is considered in this dataset,making this dataset more comprehensive and representative than the existing ground cloud images dataset.Training on the this cloud image dataset,and optimizing the trained model,compared with traditional methods,the results show that the model has better classification accuracy,fewer parameters,and higher training effi-ciency.Evaluation of a large number of experiments shows that the proposed CloudNet model can achieve good classification results in meteorological cloud classification(2)For the problem of cloud segmentation,after evaluating 12 typical semantic segmen-tation methods,we propose a semantic segmentation neural network for cloud segmentation.It learns the features between the source and target domains in an end-to-end manner.These fea-tures can solve the problem of the serious lack of labels observed in cloud image data.The model was further evaluated on the SWIMSEG dataset,using Mean Intersection over Union(MIoU),recall,F-score,and accuracy evaluation matrix.The scores of these matrices are 86%,97%,92%and 96%,which proves that the model has excellent efficiency and robustness.The most important thing is that the trained model can also achieve excellent results on the other two data sets(BENCHMARK,CCSN).Finally,the segmentation results are compared by visual method The results of segmentation are basically consistent with the results of expert judgment,which can meet the needs of meteorological operations.