Multi-head Attention and Multi-scale Feature Fusion for Remote Sensing Image Classification

Remote sensing image classification plays important role in land use and land cover monitoring applications. However, remote sensing images contain multi-scale objects which are challenging for many previous methods. Besides, high semantic features are further needed to handle the large intra-class variances and subtle inter-class difference. Different from many previous methods that design complicated structure to learn informative features from remote sensing images, a plugin multi-head attention and multi-scale feature fusion modules are designed to learn discriminate features with better semantic understanding of the remote sensing image. experiments are conducted on two widely used datasets of UC-Merced and AID. Comparisons with some relevant state methods evaluate the superiority of the proposed scheme.