Sentinel-5P Single-Image Super-Resolution Toolbox.
Code for "Model-Based Super-Resolution for Sentinel-5P Data" paper available at https://ieeexplore.ieee.org/document/10499875?source=authoralert.
All requirements for the environment used to run the codes are avilable in "requirements.txt" file.
The main file is "Main_SR_RR_Benchmark.py" from which it is possible to choose the configuration of the image to test the algorithms on. The following algorithms are tested: bi-cubic interpolation ("Cubic"), non-blind deconvolution solved with CGA and matching filters ("CGA_match"), non-blind deconvolution solved with CGA and no-matching filters ("CGA_nomatch"), SRCNN network ("SRCNN"), PAN network ("PAN"), HAT network ("HAT"), S5Net trained with matching filters ("S5Net_match"), S5Net trained with no-matching filters ("S5Net_nomatch") and S5Net without transposed convolution and bi-cubic interpolation ("S5Net_cubic").
All utility script are in the directory /scripts.
In /trained_models all pre-trained models can be found: in /S5Net both "S5Net_match" and "S5Net_nomatch" cases for each image and protocol, in /S5Net_cubic "S5Net_cubic" can be found for each image and protocol, and in /SOTA all state-of-the-art models, in particular "SRCNN" in /SRCNN, "PAN" in /PAN, and "HAT" in /HAT.
Once tested all the algorithms, if results == True, save_csv.py is run and the quality indices are saved in /results as .csv files and the super-resoluted images as .mat files. Q, ERGAS, sCC and PSNR are saved for RR protocol and BRISQUE for FR protocol.