FSBNet: A Classifying Framework of Disaster Scene for Volcanic Lithology Through Deep-Learning Models

Volcanic lithology classification is one of the fundamental tasks in remote sensing investigation of disaster scenes, and deep learning can improve the accuracy and efficiency of classification. Here, a new focused squeeze-and-excitation (SE) attention and bilinear interpolation network (FSBNet) is proposed for volcanic lithology scene classification from remote sensing image. Specifically, we first visualize and recalculate the weights of volcanic lithology features in different channels using SE attention to enhance the network’s sensitivity for extracting key semantic features in disaster scenes. Besides, the detailed features of volcanic lithology in remote sensing images are located and reconstructed bilinear interpolation and upsampling. And then the features are aggregated into the feature map by a 1×1 convolutional layer in the prediction network. Finally, we perform extensive experiments on CMRD dataset from the module, backbone, SE attention position, and ratio aspects. Experimental results show that the proposed FSBNet framework achieves 94.66% accuracy of volcanic lithology scene classification on the validation set compared with three traditional machine-learning methods and assists the remote sensing investigation of volcanic disaster scenes.