search

OCHA-DAP / pa-anticipatory-action

Code and documentation for analytical work on OCHA Anticipatory Action pilots.
GNU General Public License v3.0
14 stars 0 forks source link

Feature/mwi weighted average #238

Closed Tinkaa closed 2 years ago

Tinkaa commented 2 years ago

First attempt to weighted-averaging the ecmwf forecasts

@turnerm created functionality to compute the weighted average. I also added some functionality to divide the raster cells into smaller units with the resolution arg. The weighted average is the same as using resolution with the limit going to zero.

The code is not beautiful, but it seems to work fine. The code with weighted_average=True is veeery slow (approx 3 mins per date) so that is gonna take about 12 hours to compute..

Also tested for few dates the difference between using resolution=0.1 and weighted_average=True and there is some difference but it is small.

I computed the stats for all dates with resolution=0.1 and retrieved the confusion matrices for those, see below. The bad news is that those are even worse than using only the centres of the cells with the "rounded method" (which is already worse than only using the centres of the unrounded method)

Could you check if the code makes sense and any thoughts on what else to do?

afbeelding

caldwellst commented 2 years ago

Hi @Tinkaa, nothing I think too critical in the code that I noticed from scanning through. From the results above, I think it is not worthwhile to switch to this method. Did you originally try using all touching raster cells? Or maybe all touching raster cells where a certain % of the cell is overlapping with the southern region (to avoid where just a tiny sliver/tip is covered)? That's the only other thing I can think of to try that might be useful.

Tinkaa commented 2 years ago

Attempted to clean up as much as possible. Will merge