# Resampling assessment

I'm trying to figure out which of the resampling methods is best in my job. Comparing the original image vs. bilinear, bicubic, bicubic spline, lanczos. For now I have considered metrics like MSE RME SSIM and UQI, but I read in some articles that the SRE (signal to reconstruction error ratio) is also valid for evaluation. Should I consider it? If yes, how is it implemented on python?

Finally, the mean square error seems too high.

                                MSE            RMSE      PSNR       SSIM         UQI

NORMAL vs BILINEAR        1842447.8454      1357.3680   21.4111     0.04        0.8548

NORMAL vs CUBIC           1853309.4604      1361.3631   21.3855     0.05        0.8541

NORMAL vs CUBICA SPLINE   1834833.0737      1354.5601   21.4290     0.03        0.8552

NORMAL vs LANCZOS         1855455.1160      1362.1509   21.3892     0.05        0.8540

• Like you ask this, it's a bit open-ended. "Should I consider it?": well, yes, if your application benefits from that. But you tell us zero about your application – so there's no "best" resampling method! Notice that "best" requires you, as the user of a method, to define what you consider to be "goodness". Can't do that for you! – Marcus Müller May 18 '20 at 18:37
• Sorry, my intention is to improve the resolution of some images from 20 m to 10 m, using different methods of resampling on Sentinel 2 satellite images. – vins_26 May 19 '20 at 8:17
• doubling the resolution can't universally work (otherwise, the resolution would be twice as high to begin with), but only with some info about the image content that you have. So, I'd propose you design your metric to measure how well that worked specifically on wisely selected features that you can see at 10m resolution, but not at 20m - none of the metrics you bring up here seem to make much sense for your specific use case. – Marcus Müller May 19 '20 at 8:20
• Thanks for your reply, I'll see what I can do. – vins_26 May 19 '20 at 8:37