Multispectral satellite super-resolution must improve spatial detail without distorting spectral signatures. We propose a physics-informed dual-stream network for Sentinel-2 bands acquired at 10m and 20m. A Gaussian Mixture Model (GMM) computed from joint spectral-spatial descriptors produces soft cluster probabilities that parameterise a differentiable scientific calculator with constrained interpolation weights, yielding a content-adaptive physics baseline. A DualStreamResNet then predicts only the high-frequency residual, and the final 8band, 5m estimate is obtained by additive fusion of the two branches. Trained on 21,573 SEN2VENμS triplets with a joint $L1$ and Spectral Angle Mapper (SAM) loss and Optuna-tuned hyperparameters, the model reaches 37.65 dB PSNR and 1.37° SAM, indicating that physics-informed conditioning can sharpen spatial structure while preserving spectral consistency.
A Physics-Informed Dual-Stream Neural Network for Multispectral Satellite Image Super-Resolution Using the SEN2VENμS Dataset / Marino, A., Bottino, A., Vocca, V., Cuomo, S.. - (2026), pp. 298-302. (6th International Conference on Machine Learning and Intelligent Systems Engineering, MLISE 2026 ita 2026) [10.1109/MLISE70044.2026.11607549].
A Physics-Informed Dual-Stream Neural Network for Multispectral Satellite Image Super-Resolution Using the SEN2VENμS Dataset
Marino A.;Bottino A.;Vocca V.;Cuomo S.
2026
Abstract
Multispectral satellite super-resolution must improve spatial detail without distorting spectral signatures. We propose a physics-informed dual-stream network for Sentinel-2 bands acquired at 10m and 20m. A Gaussian Mixture Model (GMM) computed from joint spectral-spatial descriptors produces soft cluster probabilities that parameterise a differentiable scientific calculator with constrained interpolation weights, yielding a content-adaptive physics baseline. A DualStreamResNet then predicts only the high-frequency residual, and the final 8band, 5m estimate is obtained by additive fusion of the two branches. Trained on 21,573 SEN2VENμS triplets with a joint $L1$ and Spectral Angle Mapper (SAM) loss and Optuna-tuned hyperparameters, the model reaches 37.65 dB PSNR and 1.37° SAM, indicating that physics-informed conditioning can sharpen spatial structure while preserving spectral consistency.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.


