Albedo satellite method

[peterdudfield]

Remove background satellite image

Detailed Description

Remove background of satellite image by looking at the last 30 days, and removing the mean/min/max or max value. This should be time of date depends.

Context

Good to remove general background and normalise the image

Possible Implementation

• as described above
• create 30 day image, maybe mean, min and max of satellite images, and feed into model. Could also take 60, 90, ...360 image stats

[JackKelly]

Interesting idea!

Some discussion of related ideas:

• Try predicting *just* clouds (not background) by subtracting 'non-cloudy' image from actual image predict_pv_yield#17
• Use only clouds for training satflow#6
• Simulated Optical Flow using real clouds satflow#88

[jacobbieker]

Yeah, I think it'd be interesting! We could try using the cloud masks to subtract all non-cloudy pixels as well. The cloud masks for RSS only goes back till 2018, but could work as another way of doing it.

They have the issue of I don't know at what threshold a pixel is considered 'cloudy', but should be useful

[jacobbieker]

Another similar issue is openclimatefix/satflow#100 as the Weather4cast data seems to be the optimum cloud product from EUMETSAT, as well as some of the other derived data products