I am tackling a deep learning project and I need satellite images such as this one. Given I have a set of known coordinates, how and where can I get images of this resolution? Trying through EO Browser the best I get is with bilinear/bicubic upsampling and it is still blurry and not usable. The same with sentinelhub-py. Any advice on improving resolution of Sentinel products will be most welcome. My last resort will be super-resolution though I don’t think I will get much with it. Please note that I am not looking for commercial products (Pleiades etc).
The data you are after (looking at link you’ve posted) is very high resolution, and typically comes from orthophoto imagery (from planes), with resolution 50cm/px (or better). On EOBrowser I assume you have been looking at Sentinel-2 imagery, originally with 10 m/px resolution or worse (some bands are at 20m/px and a few at 60m/px).
Unfortunately, commercial imagery is the only with 50cm/px or better resolution. Your next possible solution are super-resolution (SR) algorithms on Sentinel-2 imagery (see for instance our blog post on multi-temporal SR).
If you want to stick with (fast and simple) upscaling like nearest neighbor, bilinear or bicubic that our service provides, our best results were with bilinear. Although the images look blurry, they feed well into the models we’ve trained, and we were able to obtain marginally better results than when using bicubic. The second option is to extend your DL architecture with upscaling (e.g. additional convolutional layers), training the model to X-m resolution (instead on original Sentinel-2 10m).
Thank you for this excellent advice. You cleared some really fuzzy options for me and showed me the potential of Sentinel imagery and of course your great services. Carry on the great work!