MPFASS-Net: A Mamba Progressive Feature Aggregation Network With Self-Supervised for Remote Sensing Image Scene Classification

Focusing on the semantic gap problem of “large intraclass differences and small interclass differences” caused by the complex background and scale variations of interest targets in remote sensing image (RSI), a novel Mamba progressive feature aggregation network with self-supervised (MPFASS-Net) is designed, aiming to mine the multiscale global contextual information of RSI and capture its multigranularity semantic information, thereby further improving the accuracy of RSI scene classification. First, ResNet50 is adopted as the backbone network for extracting the multilayer raw features of RSI. Second, a cross-layer feature aggregation mechanism combining dual-cross attention (DCA) and Mamba is designed, and a hierarchical feature fusion module is constructed to obtain more discriminative multigranularity features, aiming to alleviate the effects of target scale variations in RSI. Third, two self-supervised functions based on class-center vectors are constructed to alleviate the semantic gap problem in RSI scene classification. Finally, by combining the ideas of deep supervision and classifier ensemble, a new decision-level data fusion module is further designed that can explore the complementarity of hierarchical fusion features, thereby obtaining an ensemble classifier with stronger generalization ability. Comparative experimental results based on three datasets, i.e., AID, NWPU, and UCM, show that the classification accuracies of the designed MPFASS-Net model are 97.62%, 94.93%, and 99.87%, respectively, and are superior to other state-of-the-art methods.