HACloudNet: A Ground-Based Cloud Image Classification Network Guided by Height-Driven Attention

Abstract

In recent years, more and more attention has been paid to the automatic observation methods of ground-based cloud images. As it is related to real-time local weather forecasting, identifying the cloud type is always one of the basic observation items. However, due to the inability to extract subtle differences between classes, most of the existing automatic classification methods are not able to effectively recognize the cloud types defined by the World Meteorological Organization. Considering cloud images under ground-based scene have their own distinct characteristics, the proposed network architecture, called HACloudNet, exploits the informative features or classes selectively according to the vertical position of a pixel by introducing attention mechanism. We select ResNet18 as backbone network, adapt its structure to cloud classification, and combine it with the Height-driven Attention Layer, called HALayer, to guide the network to select more important features. Experiments on our ground-based scene dataset show that our method can significantly improve the performance of the backbone network. In particular, the accuracy of hard-to-classify samples has been obviously elevated. Comparison experiments show that our method is superior to the existing deep learning based cloud image classification methods without additional computational burden. It demonstrates that our method is more suitable for cloud image classification in real scenes.