Multimodal remote sensing classification benefits from complementary spectral, structural, and elevation information. Convolutional neural networks (CNNs) have limited long-range modeling, whereas Transformers incur quadratic computational and memory costs. Although Mamba enables linear-complexity sequence modeling, multimodal Mamba methods usually fuse modalities before selective scanning, limiting the direct use of auxiliary information in parameter generation. We propose Cross-Modal Selective Scan (CM-S6), which uses lightweight projections to condition selective state-space parameters Δ, B, and C on auxiliary information. These parameters regulate state retention, information writing, and state readout, respectively. With input injection, output calibration, and five tanh-bounded learnable coefficients, CM-S6 forms an input-parameter-output fusion mechanism without increasing backbone width or depth. Across ten independent runs, CM-S6 achieves OA values of 92.11 ± 0.91%, 81.04 ± 1.94%, and 92.77 ± 0.94% on Houston 2013, Augsburg, and MUUFL Gulfport, respectively, using fewer than 1 M parameters, corresponding to 1.9–2.9% of the high-capacity compared baselines. Paired tests detected no significant OA difference from MCAMamba on Houston 2013 or from HLMamba and PICNet on Augsburg, while CM-S6 shows significant OA gains over MSFMamba on Augsburg and MUUFL Gulfport and over DSFormer on MUUFL Gulfport. CM-S6(L) retains parameter-level modulation, reduces model size to approximately 0.19 M parameters, and increases throughput by 3.64–3.94×, with a dataset-dependent accuracy trade-off. CM-S6 offers an alternative for multimodal fusion under resource-constrained conditions.

Multimodal remote sensing classification benefits from complementary spectral, structural, and elevation information. Convolutional neural networks (CNNs) have limited long-range modeling, whereas Transformers incur quadratic computational and memory costs. Although Mamba enables linear-complexity sequence modeling, multimodal Mamba methods usually fuse modalities before selective scanning, limiting the direct use of auxiliary information in parameter generation. We propose Cross-Modal Selective Scan (CM-S6), which uses lightweight projections to condition selective state-space parameters Δ, B, and C on auxiliary information. These parameters regulate state retention, information writing, and state readout, respectively. With input injection, output calibration, and five bounded learnable coefficients, CM-S6 forms an input-parameter-output fusion mechanism without increasing backbone width or depth. Across ten independent runs, CM-S6 achieves OA values of 96.16±0.18%, 91.91±0.49%, and 82.18±0.46% on Houston 2013, Augsburg, and MUUFL Gulfport, respectively, using fewer than 1 M parameters-1.8-2.9% of the strongest compared baselines. Paired tests detected no significant OA difference from the strongest baseline on any dataset, while CM-S6 showed significant OA gains over PICNet and MCAMamba/MTMixer on the respective datasets. CM-S6(L) retains parameter-level modulation, reduces model size to approximately 0.19 M parameters, and increases throughput by 3.86-5.82×, with a dataset-dependent accuracy trade-off. CM-S6 offers an alternative for multimodal fusion under resource constraints.