VIS-NIR-70K: A Dataset for Nighttime Satellite Missing-Band Information Completion

The temporal interval between any two samples was ensured to be greater than one hour to reduce temporal correlation. The spatial resolution of all original observation data is 6001 × 6001 pixels. After normalizing each band to the range of 0–255, the data were converted from .nc format to .png image format to facilitate subsequent model training and visual analysis.Considering that VIS-NIR bands depend on solar radiation, their observation information is significantly weakened or even missing under nighttime or low-illumination conditions.

Therefore, this study first used the solar zenith angle (SOZ) to screen valid regions from the original samples. By setting a threshold to remove low-illumination regions with SOZ > 85, only valid illuminated regions with SOZ ≤ 85° were retained, thereby ensuring a reliable cross-modal correspondence between MIBT and VIS-NIR observations. On this basis, to reduce the storage and computational overhead caused by the original full-disk images and improve sample utilization, a non-overlapping cropping strategy was further adopted to divide the valid regions into 512 × 512 pixel image patches, thereby preserving cloud structures, texture details, and spatial contextual information.In terms of input condition design, this study selected nighttime stably observable tbb 7, tbb 8–10, and tbb 11–16 as MIBT conditional inputs, and introduced SAZ and the land–sea mask as auxiliary information to characterize differences in observation geometry and the influence of underlying surface types.

Since SOZ, SOA, and SAA have relatively weak relationships with nighttime VIS-NIR imaging, they were not used as model input conditions. The final VIS-NIR-70K dataset contains 62,290 paired training samples and 5,446 paired test samples, providing a high-quality data foundation for nighttime satellite VIS-NIR missing-band completion and the evaluation of cross-modal multiband generative models.