Xiandai Cui, Li Zhang
Accurate classification of multimodal remote-sensing imagery is critical for land-cover mapping, yet it poses a significant challenge: effectively fusing heterogeneous data (like hyperspectral, LiDAR, SAR) to leverage their complementary strengths, such as spectral signatures, elevation, and geometric characteristics, while overcoming the computational bottlenecks of existing models. While Vision Transformers excel at global-context modelling, their quadratic complexity hinders the efficient processing of high-resolution data; conversely, Mamba achieves linear complexity, but its unidirectional scan limits comprehensive spatial context integration. To address these dual limitations, we propose MTMixer, a unified encoder-decoder network that integrates Mamba and Transformer for multimodal remote sensing feature learning. At its core, a lightweight Mamba-Transformer Mixer module interleaves selective state-space blocks with self-attention layers, leveraging Mamba for efficient modelling with linear complexity and complementing it with Transformer’s global self-attention to capture comprehensive spatial relationships. A modality-agnostic alignment layer projects heterogeneous inputs into a shared latent space, enabling seamless fusion, while a symmetric encoder-decoder with skip connections preserves fine-grained boundaries. Extensive experiments on the multimodal Muufl, Houston University, and Augsburg datasets demonstrate highly competitive performance against CNN, Transformer, and Mamba baselines, achieving overall accuracies of 96.36%, 99.73%, and 97.31%, respectively. Cross-domain evaluations on Indian Pines (HSI) and Flevoland (PolSAR) further confirm its strong transferability, highlighting the framework’s efficacy and generality across diverse remote-sensing tasks.