Station Assimilation

ROI-based assimilation of station snow depth and SWE observations.

Overview

openAMUNDSEN-DA supports ROI-based assimilation of in situ station snow observations:

• station_hs for snow depth
• station_swe for snow water equivalent
• each active station contributes one likelihood term,
• stations with lower uncertainty influence the weights more strongly,
• stations with higher uncertainty influence the weights more weakly,
• all active stations on one assimilation date are combined into one ROI-level member weighting.

Supported Variables

The following assimilation-event variables are supported:

data_assimilation:
  assimilation_events:
    - date: 2023-01-10
      variable: station_hs
    - date: 2023-03-09
      variable: station_swe

Required Inputs

Station observation CSVs

Station observations live under:

obs/stations/<station_id>.csv

Expected columns:

• time
• snow_depth
• swe

Typical layout:

time,snow_depth,swe
2022-10-01 00:00:00,0.10,15.1
2022-10-01 01:00:00,0.10,14.8
...

Notes:

• snow_depth is expected in m
• swe is expected in mm
• a station file may exist even if only one variable is later assimilated
• missing values are allowed, but the requested variable must have either a unique nearest finite, nonnegative value within half the model timestep or one accepted symmetric interpolation for the station to become active

For a station event, the assimilation timestamp is the event date at the active step’s start time. Naive timestamps are interpreted in the setup timezone. Exactly two equidistant values that lie on opposite sides of the model time are averaged when both are inside the half-timestep window and their timestamps are no more than 24 hours apart. The mean is treated as an interpolated scalar at the model time without reducing its observation uncertainty. Duplicate, same-side or larger ties remain invalid, as does a value farther than half a timestep away. Model point output must exist at the exact model-clock timestamp. The matcher uses same-ID metadata rows whose use_for_da role is enabled; station IDs remain strings so leading zeros are preserved.

Station DA metadata

Station DA metadata live in:

obs/stations/stations_da_metadata.csv

Expected columns:

• station_id
• station_uncertainty_pct
• hs_sigma_abs_min
• swe_sigma_abs_min

Optional map metadata columns:

• id and name
• x and y in the setup CRS
• alt
• use_for_da and use_for_benchmark

For subdomain preparation, station_id, x and y in this table are also a complete station-selection contract. When the legacy stations_snow_depth.csv table exists it remains authoritative; otherwise the preparer selects same-ID observation series from these coordinates and disables both roles for stations that lie only in the configured station buffer.

Example:

station_id,station_uncertainty_pct,hs_sigma_abs_min,swe_sigma_abs_min
proviantdepot,20,0.05,15
latschbloder,80,0.20,

Notes:

• station_uncertainty_pct still falls back to the project-level default when the metadata value is empty.
• hs_sigma_abs_min is required for every active station_hs station.
• swe_sigma_abs_min is required for every active station_swe station.
• Missing required absolute sigma metadata is a hard configuration error.
• Before propagation, every station enabled for DA or benchmarking must have a same-ID observation CSV and openAMUNDSEN time-series output point. Matching is case-insensitive while the configured IDs are preserved. Stations disabled for both roles are exempt.
• data_assimilation.assimilation_events is final. openAMUNDSEN-DA validates its time support but does not select, substitute or remove station events. Generate different leaf schedules before execution when local support differs.
• Classified project-map station markers require station_id, x, y, use_for_da and use_for_benchmark. A benchmark-enabled station is drawn as a holdout and must not also be DA-active.

Configuration

Project-level station DA settings live under:

data_assimilation:
  station:
    default_station_uncertainty_pct: 25
    min_station_uncertainty_pct: 10
    single_station_factor: 2.0

Meaning of the keys:

• default_station_uncertainty_pct: fallback when station_uncertainty_pct is empty in metadata
• min_station_uncertainty_pct: lower bound to avoid overconfident station influence
• single_station_factor: extra sigma inflation when only one active station is available for a date

The absolute sigma floor is configured per station in stations_da_metadata.csv:

• hs_sigma_abs_min in m for station_hs
• swe_sigma_abs_min in mm for station_swe

Effective Station Uncertainty

The station uncertainty used in DA is an effective uncertainty, not only sensor error.

It can absorb:

• instrument uncertainty
• local station quality and maintenance confidence
• exposure, slope, aspect, wind redistribution and vegetation around the station
• flat-field or sheltered-location effects
• broad point-to-ROI representativeness concerns

Interpretation:

• lower station uncertainty -> narrower likelihood -> stronger update
• higher station uncertainty -> wider likelihood -> weaker update

Because percentage-only scaling can become unrealistically strict near zero snow, the implementation always combines the relative term with the station-wise absolute sigma floor from metadata.

Likelihood

For one assimilation date, each active station contributes a Gaussian log-likelihood term for every ensemble member.

Definitions:

• y_s = observed station value
• H_m,s = modeled point value at the same station for member m
• sigma_s = effective station sigma for station s

Residual:

residual_m,s = H_m,s - y_s

Station contribution:

log L_m,s ∝ -0.5 * (residual_m,s / sigma_s)^2

ROI-level member likelihood:

log L_m,total = sum over all active stations s of log L_m,s

The normalized particle-filter weights are then computed from these summed log-likelihoods.

Single-Station Handling

If only one active station is available on a DA date, the observation is still used, but more cautiously.

Version 1 policy:

• write a warning
• inflate the station sigma by single_station_factor

Effective single-station sigma:

sigma_eff = sigma_base * single_station_factor

This keeps a single station from speaking too loudly at ROI scale while still allowing it to contribute.

What Counts as an Active Station

A station is active on a DA date when:

• the station belongs to the ROI / setup station set,
• the required observation variable exists,
• the observation value is finite,
• the observation value is nonnegative,
• a unique nearest timestamp or accepted symmetric interpolation can be resolved,
• model point output exists for the station for all members.

Stations failing these checks are skipped and logged.

Outputs and Diagnostics

For station DA dates, the pipeline writes:

• weights_station_hs_YYYYMMDD.csv or weights_station_swe_YYYYMMDD.csv
• station_diagnostics_station_hs_YYYYMMDD.csv or station_diagnostics_station_swe_YYYYMMDD.csv
• one diagnostic PNG per station-DA date

The diagnostics CSV contains station-level details used for the update, such as:

• station_id
• member_id
• obs_value
• model_value
• sigma
• residual terms
• final member weights

The fraction time-series plot also marks station DA dates with vertical lines:

• HS for station_hs
• SWE for station_swe

Practical Interpretation

The station DA update does not force the ensemble mean directly to the station observation.

Instead, it reweights the members that already exist.

That means:

• if the ensemble already contains members close to the station observation, the posterior mean can move strongly toward the station
• if no ensemble member is close to the station observation, the DA can only move toward the best available members

This is standard particle-filter behavior and is especially important for strong, low-uncertainty station updates.

Related Documentation

• Configuration Reference
• Observation Processing
• Running the model
• Data Assimilation Methods