Sev2Mod Dataset

Sev2Mod dataset has been constructed by utilizing two separate instruments on board of different satellites. Importantly, both instruments operated at different spatial resolutions. Low-resolution imagery was obtained from a geostationary satellite, providing measurements every five minutes. In contrast, high-resolution imagery was obtained from a sun-synchronous polar orbiting satellite which is available only twice a day. Sev2Mod represents the task of increasing image resolution while main-
taining data availability, producing high-resolution geostationary images on a quasi-continuous basis.

In detail, input pairs were collected from the Spinning Enhanced Visible and InfraRed Imager (SEVIRI) onboard the geostationary MSG satellite, positioned at the equator at 9.5◦ East. Specifically, SEVIRI bands VIS006 (measuring around 0.6 μm) and VIS008 (measuring around 0.8 μm) served as LRMS data, while SEVIRI band HRVIS (measuring between 0.6 and 0.9 μm) provided PAN data. Output pairs were collected from the Moderate Resolution Imaging Spectroradiometer (MODIS) onboard the polar orbiting EOS-Terra satellite. MODIS bands 1 and 2 served as HRMS counterparts, operating in a similar spectral range as the SEVIRI VIS006 and VIS008 bands, re-
spectively. Over southern Europe and northern Africa, our study area, the ground sampling distance of the SEVIRI LRMS and PAN images is around 3 × 3 and 1 × 1 km, respectively. The sampling distance of the MODIS HRMS images is 250 × 250 m. To construct the Sev2Mod dataset, image pairs taken from the same area and approximately the same time (± 30 seconds) and area were collected from both satellite instruments. MODIS images were resampled using Satpy to spatially overlap with the SEVIRI grid while their native spatial resolution was approximately conserved. Subsequently, the generated tiles were cropped by a sliding window with small overlap. The resulted dimension are 16×16 pixels for LRMS, 48×48 for PAN and 192×192 for HRMS. Hence, the dataset resulted in a ×12 increase in resolution in each direction between the input and target images. Because SEVIRI and MODIS make observations with different viewing angles, spectral channels and slightly different acquisition times, and because their respective calibrations and noise levels are not identical, Sev2Mod includes inconsistencies between source and target images which one might typically encounter in advanced pansharpening applications. Hence, the performance on this dataset is more indicative of real-world performance than artificially downscaled benchmarks.

Acknowledgement:
Contains transformed Image Data excluding numerical values from EUMETSAT Rapid Scan High Rate SEVIRI Level 1.5 2009. Contains transformed Image Data, reprojected, for NASA MODIS Terra 1999