Improve sampling of tiles

[jacobbieker]

For other nowcasting applications, such as precipitation, the training data is usually sampled so that rainfall exists or is above some threshold in every, or nearly every input sample. This ensures that, while most the time there is no rainfall in a given area, the model will learn what happens when there is. For forecasting clouds and cloud movement, something similar should probably be done.

[jacobbieker]

One idea for sampling is to focus on tiles where clouds appear from nothing, instead of just moving across. These are rarer, and would be something where optical flow methods would have a very difficult time, but ML models might be able to predict?

[jacobbieker]

Also, we could make a sampler that shows whatever fraction of different topographic/time of year/etc. inputs to the model, so that the model sees a wide variety of different topologies and evenly throughout the year.