MESA: Text-Driven Terrain Generation Using Latent Diffusion and Global Copernicus Data

Image super-resolution (SR) is essential for overcoming sensor hardware limitations, yet most existing techniques primarily focus on resolution enhancement without effectively managing error propagation during intermediate stages. The Mamba model has shown promise by representing the reconstruction process as a sequence of states, enabling intermediate-level control. However, its fixed linear mapper, with a limited receptive field, constrains its ability to capture fine-grained details. To address these limitations, we introduce MambaX, a novel spectrum-state sequence model that maps spectral bands into a latent state space while dynamically learning control parameters for SR tasks. MambaX advances beyond existing approaches in three key aspects: 1) It employs nonlinear dynamic learning to more effectively approximate control matrix parameters in the latent space. 2)It introduces a cross-state control framework, enabling a unified approach to both single-image super-resolution (SISR) and multimodal fusion for enhanced flexibility. 3) It leverages transitional learning to mitigate degradation effects in remote sensing images. Extensive experiments demonstrate that MambaX outperforms state-of-the-art models in both SISR and multimodal fusion-based SR tasks, highlighting its capability for advanced spectral modeling across diverse dimensions and modalities. To ensure reproducibility and foster further research, we will publicly release the datasets and code used in this study.