Hierarchical Attention-based Fully Convolutional Network for Satellite Cloud Classification and Detection

Cloud classification and detection of satellite image are crucial for meteorological forecast. In this paper, we propose the hierarchical attention-based fully convolutional network which employs low-level attention information to improve classification and detection accuracy for satellite cloud images. Our hierarchical attention architecture combines low-level and high-level information of convolutional neural network and considers spatial neighborhood information based on the receptive field of convolution kernel. Experimental results on satellite cloud dataset indicate that our hierarchical attention-based fully convolutional network outperforms the other methods on cloud classification and detection.