MMC-Net: Mamba-based multi-level class-aware network for panchromatic and multispectral fusion classification

Recent advances in remote sensing (RS) have improved panchromatic (PAN) and multispectral (MS) data acquisition, but challenges remain in multi-source fusion classification due to limitations in non-redundant feature extraction (i.e., effectively preserving complementary information while suppressing source-specific redundancies) and cross-source alignment (i.e., accurately matching spatial–spectral correspondences between PAN and MS data). The emerging state space model (SSM)-based Mamba model, with its global modeling and linear complexity, shows potential for complex land cover classification tasks. To this end, we propose MMC-Net, a novel Mamba-based multi-level class-aware network for PAN and MS fusion classification, which integrates a progressive fusion feature learning (PFFL) module and a multi-branch class-aware classification (MBCAC) module. The PFFL module employs dual-path spatial–spectral Mamba (S 2 Mamba) to extract discriminative features while suppressing redundancy through the feature reconstruction module (FRM), along with fusion Mamba (FMamba) to enable cross-source interaction via channel exchange and Fusion SSM. The MBCAC module enhances classification diversity through spatial-wise softmax and multi-branch learning, while its parameter-sharing mechanism across levels captures hierarchical discriminative patterns. Validated on three PAN-MS and one HSI-LiDAR datasets, MMC-Net achieves state-of-the-art performance and demonstrates robust, source-agnostic generalization across heterogeneous modalities. The code is available at https://github.com/Xidian-AIGroup190726/MMCNet .