Improved RepVGG Ground-Based Cloud Image Classification with Attention Convolution

Abstract. Clouds greatly impact the earth's radiation prediction, hydrological cycle, and climate change. Accurate automatic recognition of cloud shape based on ground-based cloud image is helpful to analyze solar irradiance, water vapor content, and atmospheric motion, and then predict photovoltaic power, weather trends, and severe weather changes. However, the appearance of clouds is changeable and diverse, and its classification is still challenging. In recent years, convolution neural network(CNN) has made great achievements in ground-based cloud image classification. However, traditional CNN has a poor ability to associate long-distance clouds, so extracting the global features of cloud images is difficult. Therefore, a ground-based cloud image classification method based on improved convolution neural network RepVGG and attention mechanism is proposed in this paper. Firstly, the proposed method increases the RepVGG residual branch and obtains more local detail features of cloud images through small convolution kernels. Secondly, an improved channel attention module is embedded after the residual branch fusion, which can effectively extract the global features of the cloud images. Finally, the linear classifier is used to classify the ground cloud images. In addition, the warm-up method is introduced to optimize the learning rate in the training stage of the proposed method, and it is lightweight in the inference stage, which can avoid over-fitting and accelerate the convergence speed of the model. The proposed method in this paper is evaluated on MGCD and GRSCD ground-based cloud image datasets, and the experimental results show that the accuracy of this method reaches 98.15 % and 98.07 %, respectively, which exceeds other most advanced methods, and proves that CNN still has room for improvement in ground-based cloud image classification task.