Review Kenneth Nowak
This method and documentation is much more parsimonious. The documentation wasn’t in the requested place, but I was able to located it. From what was provided, the method breaks the domain into 24 sub-domains and uses a random forest to predict the target variables for each cell within that domain. Cited inputs/data are ECMWF forecasts, observations, and ‘El Nino”. Although the documentation was light and I did have some questions, I’d have a hard time giving a “fail” rating.
Questions:
- I’m unclear if/how El Nino data are being used. The link goes to a general data page at NOAA. This is something I’d inquire on. There is some code re: SST data, but it is commented out and I’m not sure if that is/isn’t what is meant by “El Nino”. If ENSO is being used, what guided the decision to add those data as a predictor? My understanding is that NWP models are initialized/assimilate current SST data and are fairly “ENSO aware”. As such, adding ENSO as a predictor is a bit surprising (redundant?), although not totally unreasonable. Further, would be curious to know if ENSO is used in all 24 sub-domains, how were those decisions made and can comment be offered on the value the ENSO data add/don’t add in each of the 24 different sub-domains.
- Following on the 24 sub-domains, would be interested to hear more about the background/motivation for breaking the domain up – what if any references might be provided and why was it done the specific way it is implemented? Related, are the sub-domain RF models completely independent of each other? If so, does this approach risk unrealistic spatial co-variance between the 24-sub domains, or does the ECMWF forecast input sufficiently “guide” realistic spatial patterns?