Galapagos refined analysis validation data

A 2024 dataset release bringing together observations from the DARWIN station network for evaluating the refined analysis. It is a dataset with its own citation, not a supplement to a paper, which matters — the observations are useful to anyone working on the archipelago regardless of what the model did with them.

What it is for

Evaluation, and specifically process-oriented evaluation. The distinction shapes what the dataset contains.

An aggregate skill assessment asks how closely modelled values match observed ones across everything available. It is easy to compute and weak evidence, because a model can score well through compensating errors — the right seasonal rainfall total arrived at by raining too much in one regime and too little in another. Here that would be disqualifying, since the entire question is how moisture partitions between rainfall and fog interception across elevation.

So the dataset is assembled around the comparisons that discriminate: does the model reproduce the elevation gradient, does it get the seasonal transition at the right time and elevation, does the near-surface humidity structure behave like a capped marine boundary layer should. Near-surface temperature, humidity, wind, and precipitation are the variables, but the framing is which contrasts they support rather than how many station-hours are included.

What it establishes

It is the reason the refined analysis can be used to say anything about places without stations. The model’s credibility over the ungauged parts of the archipelago rests entirely on how it performs at the gauged ones, which makes this dataset the load-bearing element of everything downstream of it — the publication thread included.

Scope and limits

Station positions are generalised to elevation band and environment class, and site-level operational detail is not included. That constrains what someone could do with it — precise co-location analysis is not possible — and it preserves what the dataset is for, which is establishing whether the modelled gradients are real.

See also: the DARWIN project, Cerro Crocker, datasets generally, and hydroclimate.