LAI time series forecasting

[Yuval-Sadeh]

Feature description

When we want to do forecast in-season we don't have the full LAI profile.
Here is an approach that enable to forecast time series.

Suggested solution

This proposed solution:
https://github.com/PriorLabs/tabpfn-time-series
can potentially be used to forecast LAI patterns in a region.
we can us historical RS LAI estimates of a given region, or to use synthetic LAI from APSIM.
What do you think?

Additional context

No response

[rohansaw]

This sounds super interesting! We should do a small experiment by just creating some time series for a few regions (e.g 2019-now) and create a train-validation split from these (+ possibly add the met data as features aswell). If i understand the model correctly no finetuning is required, so we could just do some benchmarks on this and if it is sufficiently good plug it into the pipeline.

However, I would prioritize this lower for now and would first focus on improving the non-forecasted scores.

[Yuval-Sadeh]

OK
What do you mean by "non-forecasted scores"?

[rohansaw]

I was just referring to our general predictions using the LAI&Met data of the complete season instead of having a cutoff of x dates to the season end.