verify against Argo profiles

[miniufo]

I have a hindcast example, with the gridded dataset:

<xarray.Dataset> Size: 671MB
Dimensions:  (lead: 7, init: 31, lat: 941, lon: 821)
Coordinates:
  * lon      (lon) float64 7kB 99.02 99.07 99.12 99.17 ... 139.9 140.0 140.0
  * lat      (lat) float64 8kB -11.98 -11.93 -11.88 -11.83 ... 34.92 34.97 35.0
    lev      float64 8B ...
  * init     (init) datetime64[ns] 248B 2025-08-01 2025-08-02 ... 2025-08-31
  * lead     (lead) int64 56B 0 1 2 3 4 5 6
Data variables:
    temp     (lead, init, lat, lon) float32 671MB dask.array<chunksize=(1, 31, 941, 821), meta=np.ndarray>

and observations from Argo profiles in the ocean:

<xarray.Dataset> Size: 2MB
Dimensions:  (N_PROF: 217, N_LEVELS: 1011)
Coordinates:
  * N_PROF   (N_PROF) int64 2kB 0 3 18 66 88 91 ... 4383 4385 4391 4435 4439
Dimensions without coordinates: N_LEVELS
Data variables:
    temp     (N_PROF, N_LEVELS) float32 878kB ...
    lons     (N_PROF) float32 868B 125.8 130.3 128.8 128.6 ... 125.6 129.4 126.3
    lats     (N_PROF) float32 868B 21.11 6.102 18.73 27.07 ... 21.8 22.67 20.47
    time     (N_PROF) datetime64[ns] 2kB 2025-08-01T21:54:01 ... 2025-08-31T0...
    dpth     (N_PROF, N_LEVELS) float32 878kB ...
Attributes:
    long_name:       Sea temperature in-situ ITS-90 scale
    standard_name:   sea_water_temperature
    units:           degree_Celsius
    valid_min:       -2.5
    valid_max:       40.0
    C_format:        %9.3f
    FORTRAN_format:  F9.3
    resolution:      0.001

How to verify the gridded dataset against this Lagrangian type of Argo observations? The observations is not a standard spatial-temperal dataset. It is indexed with profiles and levels. time is a variable as temp, so it is not unique, as in a single time, there may be more than one profiles. This kind of verification is very common in the oceanic context. Just want to know how to do this using climpred framework.

[aaronspring]

you need to somehow get your obs to the same grid as the forecast. your forecasts miss depth.
sorry, can help one the specifics.

[miniufo]

The problem is not due to depth. I can simply drop that so we could focus only on 2D horizontal space. The peoblem is due to that observation are not indexed with time. So basically, climpred requires observations as a gridded dataset?

[aaronspring]

climpred requires observations and forecasts to have the same coordinates for spatial/depth/height dimensions, the same way as in xarray forecast - observations

[miniufo]

If I have a single point forecast as init+lead time series, and some observations randomly sampled at some times (not evenly distributed), is it possible to use climpred in this case? I don't know the design under the hood. But from the added valid_time dimension in the hindcast class, I feel this could be done with some extra efforts to align the time.

[aaronspring]

Under the hood climpred matches init+lead with obs time by exact matches. You need to ensure there are these matches in your input. So if you have random obs you might gets some matches. Use alignment="maximize".