MSCE-Net: Multi-scale Spatial and Channel Enhancing Net based on Attention for Cloud Image Classification

Abstract

The difference between cloud types is mainly present in appearance but has problems of high similarity and insignificant appearance between classes. In addition, the existing approaches separately use the channel and spatial attention machines in a stage of the network that loses the attention in multiple scales and the relationship between channel and spatial attention. Therefore, this study proposed a framework, namely multi-scale spatial and channel enhancing Net (MSCE-Net), based on attention mechanisms to enhance the feature learning ability of the network for cloud image classification. The designed multi-scale spatial and channel attentions consider information between scales and concatenate spatial and temporal attention at each scale to consider the relationship between spatial and channel attention and make the network focus on the appearance of the cloud image, which is significantly different between various clouds. The main contributions of this paper are as follows: (1) we construct a new cloud image dataset with ten categories with a total of 6000 images; (2) we generate multi-scale spatial-based and channel-based enhancing factors to enhance the spatial and channel features, respectively, at each scale; (3) we concatenate spatial and temporal attention at each scale to consider the relationship between spatial and channel attention at each scale. Experimental results show that the accuracies of the proposed framework suppress the state-of-the-art approaches and achieve 85.11%. Moreover, the ablation experiments prove the efficiency of the proposed strategies.