<p>Volcanic disaster scenes have diverse types and random distribution, displaying complex global features, local information, and sample label ambiguity in remote sensing images. The existing convolutional neural network (CNN)-based classification for remote sensing images is limited by the fixed receptive field of the convolutional kernel, which reduces the modeling ability of local feature and long-term dependencies in remote sensing images. To address this issue, a rough set attribute reduction framework for the Res-Attention_Unet network (RSARNet) used for volcanic disaster scene classification is presented in this paper. In RSARNet, the rough set attribute reduction module uses genetic algorithms to dynamically reduce decision tables and remove redundant attributes so as to better overcome the sensitivity of the network to parameter settings and dependence on sample selection. The Res-Attention_Unet module explores the multi-scale deep features of volcanic disaster scenes by focusing on global contextual information and local details. And then the fully connected layer and classifier are combined to implement the prediction of volcanic disaster scene and output of classification labels. Finally, a volcanic disaster scene (VDS) dataset was used to test the feasibility of the proposed method. Extensive experimental results show that the RSARNet method has the most significant improvement effect on volcanic disaster scene classification, with an overall accuracy of 92.54% compared to traditional machine learning methods. The findings of this paper provide new references for using remote sensing and deep learning for volcanic disaster monitoring and disaster prevention and reduction.</p>