A novel approach for JWST image fusion

Context . Hyperspectral-multispectral fusion enhances low-resolution hyperspectral images so that they match high-resolution multi-spectral data. For JWST astronomical imaging, this is challenging because the data dimensionality is extreme, the resolution varies spectrally, the sampling is limited, and there is high noise. Combining an imager and an integral field spectrometer, the MIRI midinfrared instrument on board the JWST creates opportunities for high-definition spectral mapping, but requires joint super-resolution, denoising, and deconvolution. Aims . We developed a simulation pipeline that models the full JWST instrument responses to enable supervised training on realistic synthetic datasets. This framework trains a deep-learning architecture for joint fusion, denoising, and deconvolution of the MIRI imager and Medium Resolution Spectrometer data to reconstruct spatially super-resolved hyperspectral images with full spectral fidelity. Methods . The training relied on synthetic data generated by simulating complete JWST MIRI instrument responses. We then introduced an innovative two-branch neural architecture estimating compact parametric correction functions per pixel. A spline-based branch modeled smooth spectral continua, while a high-frequency branch restored the spectral lines. Results . The method achieves stable optimization on thousands of spectral channels, enables super-resolution beyond the guiding images, and reconstructs broad continua and narrow features under high noise. We report the first deep-learning deconvolution of real JWST MIRI multispectral and hyperspectral images on the reflexion nebula NGC 7023. Conclusions . These results demonstrate the effectiveness of compact parametric fusion for astronomical multispectral and hyperspectral imaging and establish deep learning as a practical tool for JWST data fusion and deconvolution.