Tri-CoMamba: A Tri-Complementary Mamba Framework for Multisource Remote Sensing Image Classification
Year: 2026
Type: article
Abstract: The synergistic application of hyperspectral images combined with light detection and ranging (LiDAR) or synthetic aperture radar (SAR) data is crucial for improving the accuracy in multisource remote sensing joint classification. However, existing methods still suffer from limitations in long-range dependency modeling, cross-modal alignment, and the preservation of fine-grained spectral features. This study introduces the tri-complementary Mamba modules (Tri-CoMamba) framework to resolve the aforementioned limitations. The proposed network architecture is founded upon the state-space model and employs selective scanning Mamba-S6 as its core structure, integrating three complementary modules: complement-and-rectify Mamba (CoRe-Mamba), cross-frequency spectral Mamba (CF-SpecMamba), and modality-aware spatial modulation (MASM). Specifically, CoRe-Mamba mitigates feature mismatching through dual-level spatial and channel rectification, thereby enhancing semantic consistency and directional modeling capabilities. CF-SpecMamba introduces bidirectional recurrence and cross-frequency interaction attention to balance low-frequency baselines with high-frequency details in spectral modeling, to achieve comprehensive spectral feature enhancement. Furthermore, MASM utilizes modality-aware dynamic spatial modulation to highlight discriminative regions and suppress background interference, thereby optimizing the cross-modal fusion effect. The synergy of these three modules enables Tri-CoMamba to completely exploit the distinct yet supportive strengths of spatial, spectral, and modal features, all while preserving computational efficiency, which leads to the precise classification of multisource data. The effectiveness of this approach was validated using the Berlin, Trento, and Houston2018 datasets, with results demonstrating that Tri-CoMamba outperforms various representative methods, achieving overall accuracy of 78.51%, 99.80%, and 92.88%, respectively.
Authors: Zhihui Geng, Jiangtao Wang, Rui Wang
Institution: Huaibei Normal University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,800
OpenAlex ID: W7128310906