SpecSwin: A Hyperspectral Data Simulation Framework for Scalable Geospatial Modeling
Tang Sui, Songxi Yang, Qunying Huang
Accurate and scalable geospatial simulation often requires high-fidelity spectral information, which is limited by the trade-off between the high spatial but low spectral resolution of multispectral imagery and the rich spectral but lower spatial resolution of hyperspectral imagery. This paper introduces SpecSwin, a transformer-based framework for generating high-resolution hyperspectral imagery from limited multispectral observations. In our approach, five multispectral bands from California and Nevada are used as inputs to generate 224 hyperspectral bands at matching spatial resolution, enabling realistic simulation of dense spectral datasets for downstream geospatial analysis and modeling. To enhance spectral fidelity, we propose a cluster-wise reconstruction strategy that groups target bands by spectral proximity, as well as an optimized band sequencing scheme that strategically repeats and orders the input bands to maximize local inter-band interaction. Quantitative evaluation yields a PSNR of 34.13, ERGAS of 0.80, SAM of 2.66°, Q Index of 0.95, and SSIM of 0.95, confirming the potential of SpecSwin as a reliable tool for generating realistic geospatial simulation data.