question: adjusting target ldcast

[tomas van oyen]

Hi @jleinonen,

more an approach question.

Say, you are just interested in the rainfall at one particular spot (e.g. discharge at one point in a hydrological basin). Do you think it is worthwhile to adjust the target to the time series of the future precipitation (instead of the time series of the radar images)? This implies, among others, skipping the encoder part of the future precipitation.

Look forward to your thoughts on this.

Regards,

Tomas

[Jussi Leinonen]

Hi Tomas,

My first guess is that if you specialize the model specifically for one basin, you may well get better results. But in that case you probably don't even need a generative model at all, it's easier to use a simpler predictive model since you're only interested in one point.

The nice thing about generative models is that they give you the flexibility to look into spatial and temporal distribution properties of the rainfall probabilistically over a wide range of scales. But if you are interested in one point only you don't really need that flexibility.

Cheers,
Jussi

[tomas van oyen]

Hi Jussi,

I have to ponder a bit on this, and will come back to it later.

Regards,

Tomas

[tomas van oyen]

Hi Jussi,

going through the code, it appeared to me that indeed AFNONowcastNetBase could be used directly to predict the future water levels at a single point, starting off from the trained weights on the latent diffusion problem.

I am still wondering if using the decoder as in AFNONowcastNet would be helpful, but I guess I will just have to try.

I welcome any thoughts on this.

Regards,

Tomas